pmacs/docs/generated-buffer-immutability-framing.md

Generated-buffer immutability

PROPOSED — needs explicit user approval before implementation. DO NOT implement, DO NOT merge.

Revision 6 — scouted against canonical githubsucks/main @ 64883eb, merged into this branch at 76cfaac, 2026-07-29. Both SHAs were measured in this worktree at the moment of writing (git rev-parse githubsucks/main, git rev-parse HEAD), not carried forward from a briefing. Every count and every pre-image below is a reading at 76cfaac, not a constant; the base has moved four times while this document has been open and will move again.

Numbering note. The review that this revision answers asked for "revision 5". The document at the reviewed commit cab3404 was already revision 5 (written by 238fd04, closing review round 4), so this is revision 6. Recorded rather than silently renumbered, because two documents disagreeing about which revision is live is the failure this arc's cross-lane boundary exists to prevent.

Revision 6

Answers review round 5 on cab3404 — five P1, two P2. All seven confirmed. It withdraws two of revision 5's decisions, corrects one assertion revision 5 made about containment, and reclassifies six criteria.

finding	the decision
P1-1 — Stage 1 criterion 5 is impossible	Confirmed. begin_edit (src/buffer.rs:724) and apply_edit (:772) each call ensure_writable() as their first statement, so once the rope is locked an ordinary edit necessarily returns ReadOnly and no intercept runs. The named-error requirement is dropped from the ordinary path and moved behind a Rust-side lift, where the intercept genuinely is the guard. The wider error-precedence change is considered and rejected. Consequence revisions 1–5 all missed: after adoption the intercept is unreachable on the ordinary path including in the shipped precedent — terminal.lua's intercept has been dead since #178. Criterion 5 splits into (a) refused-and-unchanged and (b) refused-by-the-intercept-under-a-lift, which is where the original bite survives.
P1-2 — Diverged misses another post-mutation failure	Confirmed, and revision 5 got three of seven cases wrong, not one. export_updates_since (src/buffer.rs:1173) runs after every successful CRDT mutation, for all three op shapes. Revision 5's "delete→insert classifier" maps all three of those to Rejected, whose cleanup restores a fresh buffer to writable while the CRDT and rope disagree — reintroducing the exact harm round 4's P1-3 withdrew the four-variant fallback to prevent. Replaced by a crdt_mutated flag carried out of the routing function, which is total by construction rather than an enumeration that a future failure point can escape.
P1-3 — read_only does not contain a divergent CRDT	Confirmed; revision 5's "strongest available containment" was simply wrong. Neither outbound snapshot path reads read_only (src/daemon.rs:2563-2578, :2693-2708) — and neither can, since read_only is a statement about inbound mutation. New crdt_quarantined flag with three consumers: both export sites and queue_daemon_origin_crdt_op, the third of which the review did not name. Repair stays deferred, and the cost of quarantine (the buffer stops collaborating) is stated rather than buried.
P1-4 — AppliedThenFailed leaves later views stale	Confirmed. Revision 5 fixed the fan-out after the borrow drops and left the loop inside it broken. The review named one broadcast site; sweep F found four, and two of the three it did not name are load-bearing for this arc specifically: src/buffer.rs:1250 is the arm every NoOp takes — including Q#GB5's ensure_slot lock — and :1033 is how a locked buffer's write reaches a replica at all. All four adopt continue-and-retain-first-error. Criterion 15a gains a RecordingView after FailingView.
P1-5 — Q#GB6 omits the selection anchor	Confirmed, and it is the only one of the three coordinates that panics: Window::region (src/window.rs:472-479) clamps neither endpoint, and region_bytes (src/editor_core.rs:4184-4191) hands the result to Rope::slice, which asserts at src/rope.rs:145. #191 reproduced it. The rule is clamp-or-clear, in both clamp sites, for every affected window — and it is not invented here: src/terminal/view.rs:715-721 already solved the identical problem for TerminalSelection, collapsed_by_clamp and all. New criterion 8c. Also found: a stale anchor reaches the presence broadcast (src/presence.rs:122-123), not only the local slice.
P2-6 — four criteria have incorrect main pre-images	Confirmed, and there are six. 15, 16, 16b, 17, 18 and 21 all pass at the base. Relabelled to [mutation] / [fix-shape] with a table stating, per criterion, what the base actually does. None is weakened — each already carried its falsifying mutation.
P2-7 — a rejected generated edit can still unfold	Confirmed. A generated_preflight runs §3.4's exits 1–4 under a read-only borrow before the unfold; apply_generated_edit still re-checks all four, because the borrow is released between and Buffer stays the only authority. Dropping the unfold and narrowing the contract were both considered and rejected. New criterion 22, with a second half on the identity-protected refusal so it pins the preflight rather than one hoisted check.

Sweep F — every recovery or containment mechanism in the document, audited for failure points it does not enumerate (obligation 1). The review is right that P1-2, P1-3 and P1-4 are one class. 10 mechanisms examined, 4 broken — all four are the review's — and within them three exits the review did not name.

#	mechanism	verdict
1	Diverged classifier	BROKEN — enumerated 1 of 3 failure points; now flag-based and total
2	Diverged containment via read_only	BROKEN — 2 export paths, plus queue_daemon_origin_crdt_op (not named by the review)
3	AppliedThenFailed view notification	BROKEN — 4 broadcast sites, review named 1
4	Q#GB6's coordinate clamp	BROKEN — 3rd coordinate; and it also reaches presence, not named by the review
5	Rejected restore-entry-state	clean — exits 1–4 plus CRDT-untouched, and total once crdt_mutated is the discriminator
6	NoOp cleanup	clean — one exit (src/buffer.rs:1245)
7	generated_preflight (new)	clean by construction — evaluates the same four predicates apply_generated_edit does
8	clear_history across modes	clean — if let Some(crdt) guard covers both
9	identity_protected write refusal	clean — single exit
10	Q#GB12's revision guard going near-dead	clean — names its residual writer (a future Rust-side one)

Sweep G — every criterion's pre-image re-derived by asking what the BASE does (obligation 2). Re-derived at 76cfaac; see the table above criterion 15. 41 criteria audited across both stages, 6 mislabelled, all relabelled. The arithmetic, since a count asserted without one is what this sweep exists to catch: Stage 1 has 14 top-level entries of which 13 is a split header carrying no assertions, plus 4 lettered sub-criteria (8b, 8c, 13a, 13b) — 14 − 1 + 4 = 17. Stage 2 has 25 entries of which 10 is "Coverage, not a criterion" — 25 − 1 = 24. 17 + 24 = 41. The rule, now stated for the third time and applied by construction rather than by inspection: a criterion's pre-image is a fact about the base, established by running it there. Because the base moves, §6's labels carry the commit they were derived at.

Re-measured at 76cfaac (obligation: paste the output).

$ git rev-parse --short HEAD; git rev-parse --short githubsucks/main
76cfaac
64883eb
$ grep -rn 'describe\.buffer' builtin/ --include='*.lua' | wc -l
14
$ grep -rn bypass_intercept builtin | wc -l
21
$ grep -rn --include='*.lua' add_intercept . --exclude-dir=target | wc -l
17
$ grep -rn "is_modified()" src/ pmacs-gpu/ | wc -l
26
$ grep -rn "mark_clean()" src/ | wc -l
6
$ grep -c "dispatch_idle_for" tests/terminal_copy_mode_acceptance.rs
0

All unchanged. Verified rather than assumed: git diff --stat 300cbc4..64883eb touches only docs/ and scripts/bite — no product code — which is why the source citations below survive the base move.

Citation audit, revision 6. PR #187 moved four source files, and a sweep of every citation in this document into them found 15 stale instances across 9 distinct citations, all corrected here: editor_core.rs :1814→1836, :1843→1865, :1841-1842→1863-1864, :1853-1857→1875-1879 (×3), :2575→2597, :1945→1967; editor.rs :1436,1984,2128,2163→1472,2356,2500,2535, :3704→4129 (×2); daemon.rs :2976→3167; semantic_render.rs :1347→1716 (×2). The tell was internal: revision 6's new text cited editor_core.rs:1836 and :1865 for the same two functions that older text still cited as :1814 and :1843. A citation is a reading too, and this document had not been re-reading them.

scripts/bite now has a positive control (main @ 64883eb), so every "falsify by …" below can be executed rather than asserted: exit 0 is a real bite, exit 3 is NO CONTROL — the named tests did not pass on the working tree, or none ran — exit 1 is vacuous, exit 4 is INCONCLUSIVE (MIXED), and the swapped run is labelled OK (assertion) or OK (COMPILE). Prefer OK (assertion); treat OK (COMPILE) as a prompt to narrow the swap. Criteria whose bite is a one-line mutation of a Rust file are directly bite-able; the Lua-side ones swap the builtin/ file.

Revision 5's decisions and its answers to review round 4 remain below as history. Where its live design said "delete→insert classifier" or called read_only containment, revision 6 supersedes it explicitly.

Revision 5

Answers review round 4 on d98d0b3 — four P1, one P2. All five findings are confirmed. This revision removes one proposed public capability, makes one transaction outcome larger, and makes the CRDT fault seam mandatory rather than leaving a fallback for implementation time.

finding	the decision
P1-1 — AppliedThenFailed loses the Edit, so the binding cannot fan out a mutation before returning Err	Confirmed. AppliedThenFailed becomes { edit: Edit, error: BufferError }; apply_generated_edit returns the outcome rather than collapsing it to Result inside Buffer. After the registry borrow drops, the binding fans out Applied, NoOp, and AppliedThenFailed. The last arm then surfaces its error. That preserves the window-cache and replica-mirror invariant even when a View::on_edit failure occurs after the rope swap. The Rust return type of set_generated_contents changes with the internal transaction; its Lua name and call signature do not. New criterion 15a pins both the window and CRDT directions.
P1-2 — criterion 15 probes the edit flag while read_only masks it	Confirmed. begin_edit calls ensure_writable before it checks editing_in_progress, while AppliedThenFailed deliberately relocks. The old follow-up therefore returned ReadOnly for both the correct and broken implementations. Criterion 15 now performs the same Rust-side lift already used by criteria 4, 16b and 17 before issuing the ordinary edit; only then can its outcome distinguish the cleared flag from ConcurrentEdit.
P1-3 — the four-variant fallback maps divergence to Rejected and restores a fresh buffer to writable	Confirmed. The fallback is withdrawn. A CRDT delete-success/insert-failure must be Diverged, must leave the buffer locked, and must surface distinctly. Criterion 16c is reclassified as a crdt-only fault-injection unit test: Stage 2 extracts a private delete→insert classifier whose production closures call loro and whose test closures force delete Ok / insert Err. There is no implementation-time choice to weaken containment because staging is inconvenient.
P1-4 — identity_protected is not generated-lock provenance	Confirmed. Q#GB7 now chooses the fallback it previously named: this arc ships no unlock_generated at all. identity_protected remains, but only as a monotonic terminal-identity guard in the write direction; it is never described as proof that the generated primitive installed a lock. The proposed setter becomes a crate-private, one-way mark_identity_protected() so “set once” is enforced rather than documentary. A future wdired door stays with dired Stage 3 and must frame the provenance/ownership transition it actually needs.
P2-5 — Stage 2 still instructs implementation to build revision 3's rejected shared registry	Confirmed. The staging checklist now matches Q#GB18 exactly: compile.lua answers from slot_for_buffer, default.lua answers from search_panel_owns, and each capture site consults the other through a guarded optional predicate. No owner registers buffers with the other, and no teardown work is introduced.

Reachability recheck for the changed criteria. Criterion 15 now crosses the post-failure lock only by a Rust-side lift, so the next edit can actually reach the flag it asserts. Criterion 15a drives the same valid FailingView mutation through the production Lua finisher and observes its window and replica consumers. Criterion 16c drives the otherwise-unconstructible second-op failure through the private delete→insert classifier. Criterion 13 is deliberately structural because revision 5's decision is the absence of a capability. None passes merely by failing before the mechanism under test.

Revision 4's measurements and its answers to review round 3 remain below as history. Where its live design said “bounded unlock” or offered the four-variant CRDT fallback, revision 5 supersedes it explicitly rather than silently smoothing the reversal over.

Revision 4

Answers review round 3 on 8e032d7 — three P1, two P2 — and two sweeps the review asked for by class. It withdraws three of revision 3's own decisions. All five findings are confirmed; none is re-litigated.

finding	the decision
P1-1 — the revision predicate is wrong in two directions	Confirmed, and it is wrong in three. §3.4 is rewritten around an explicit GeneratedOutcome reported by the apply instead of a predicate inferred from revision: five variants, each with its own cleanup, tabulated. Direction A (successful no-op) is fixed by restating the invariant as a property of the buffer — a generated-locked buffer carries no history — so NoOp clears, which needs no reference to revision and is therefore mode-independent. Direction B (CRDT mid-transaction) gets its own Diverged variant that clears nothing and surfaces, because clearing would destroy the last local record of the pre-edit rope. Third direction, found while fixing the other two and not reported by the review: revision 3's unconditional relock locked a fresh buffer that was never successfully written — a mid-codepoint generated insert on a writable *scratch* returned Err and left it read-only. Rejected now restores the entry value.
P1-2 — criteria 15-16 never enter the transaction	Confirmed, and the contradiction was internal: revision 3 argued in §3.4 that pre-validation makes an invalid range cost nothing, then used an invalid range to test the post-apply cleanup. Both criteria are rewritten around a valid write that fails at the on_edit broadcast, staged with a Rust-side FailingView (pmacs::view::View is pub, src/view.rs:221; Buffer::attach_view is pub). Criterion 16 splits into relock-on-failure and no-lock-on-refusal, which §3.4 now answers differently. New 16b and 16c cover the two P1-1 directions. Rule adopted: every criterion names the exit it drives the implementation to, and that exit must be inside the mechanism under test.
P1-3 — provenance invalidates the lift-and-restore seam	Confirmed. Q#GB15's generated_lock is withdrawn. The defect is not the rule's details but that a derived fact must be maintained by every mutation of what it derives from, and set_read_only is pub. Replaced by identity_protected — a property of what the buffer is, set once by TerminalSession::open, never written by set_read_only. tests/terminal_copy_mode_acceptance.rs:578-584's lift/upgrade/restore cycle is transparent to it, so Q#GB12 and Stage 2 criterion 4 need no change. The crdt-gated terminal suite is now named explicitly in §10.
P2-4 — search registration adds a hard dependency on compile.lua	Confirmed. Revision 3 replaced an existing guarded, optional dependency (default.lua:991-994's triple check) with an unguarded call. Q#GB18 is rewritten to symmetric guarded optionality: each module answers for its own buffers and consults the other through the guard shape already in the tree. That also removes the teardown obligation revision 3's registry introduced. New Stage 2 criterion 21 pins the minimal harness.
P2-5 — fold criterion 13 contradicts its classification	Confirmed. If dired folds successfully on main, asserting false fails on main, so [fix-shape] was the wrong label. Split into 13a [main] (the behaviour change, falsifiable by revert) and 13b [mutation] (the status string, whose pre-image cannot be main because on main the call succeeds and sets no status).

Sweep D — rules spanning the v0.1/CRDT boundary or the fresh-lock/existing-lock boundary (obligation 1). The review is right that P1-1 and P1-3 are one class: a rule derived from one mechanism and applied to a second that does not share its ordering. 13 rules examined, 3 broken, 1 verified that looked at risk.

#	rule	boundary	verdict
1	cleanup keyed on revision	v0.1 / CRDT	BROKEN — P1-1
2	clear_history clears whichever history exists	v0.1 / CRDT	clean — reads both, src/buffer.rs:558-566
3	Q#GB4's O(1) clear argument	v0.1 / CRDT	clean — in CRDT mode self.undo is bypassed entirely (:1268-1272), so the "at most one entry" premise is vacuously true rather than wrong
4	Stage 2 criterion 4 asserts Err(NothingToUndo)	v0.1 / CRDT	verified, not assumed — undo_crdt_mode returns BufferError::NothingToUndo when the manager reports nothing undone (src/buffer.rs:1374-1376), so the same assertion is correct in both modes
5	§3.3 "CRDT behaviour: nothing new"	v0.1 / CRDT	corrected — true except for Diverged; §3.4 now says so
6	Q#GB10's mark_clean	neither	clean — is_modified is mode-independent
7	Q#GB6's coordinate clamp	neither	clean
8	criterion 10's coverage rule	v0.1 / CRDT	tightened — it now names which criteria are irreducibly single-configuration
9	refuse "someone else's lock" by provenance	fresh / existing lock	BROKEN — P1-3
10	relock unconditionally	fresh / existing lock	BROKEN — found here, not reported; a refusal locked a buffer nobody wrote
11	Q#GB5's empty set_generated_contents in ensure_slot	fresh / existing lock	clean only because Q#GB13 makes the buffer fresh by construction — and it is also the concrete caller that made P1-1 direction A load-bearing rather than theoretical; the dependency now runs both ways and is stated in Q#GB5
12	unlock_generated's refusal	fresh / existing lock	follows 9; fixed with it
13	Q#GB2's generated implies bypass	neither	clean

Sweep E — every criterion added in revisions 3 and 4, audited for whether the state it asserts is reachable by the operation it performs (obligation 2). This is the vacuous-criterion class, which this arc has now shipped twice — revision 2's Stage 2 criteria 3-5, then revision 3's 15-16. 19 criteria audited, 3 defects, all fixed here.

criterion	reachable by its own operation?
S1-6 (round-trip, 3 halves)	yes — and (c) is what makes (b) non-vacuous
S1-11 (disambiguated panel keys)	yes — foreign buffer first, panel opens as <2>
S1-12 (q-target not inverted)	yes — needs the two-panel sequence, which is why it is separate from 11
S1-13 (fold)	MISCLASSIFIED — P2-5; split into 13a/13b
S1-14 (structural)	n/a
S2-6 (four surfaces)	yes
S2-13 (unlock, 3 halves)	yes — the terminal half needs a real terminal, which the copy-mode suite already opens
S2-14 (refuse a foreign lock)	yes — fails on main
S2-15 (flag cleared on error)	VACUOUS — P1-2; rewritten around FailingView
S2-16 (relock on error)	VACUOUS — P1-2; rewritten and split
S2-17 (invalid range keeps history)	yes — and it is the one criterion for which exit 4 is the mechanism under test, so it stays as written
S2-18 (re-entrant refused)	yes — phase 2 of run_managed_edit releases the borrow (src/lua_bindings/mod.rs:1477-1487), so the inner call reaches the buffer and finds the flag set
S2-19 (is_generated_buffer disambiguated)	yes
S2-20 (structural)	n/a
new in rev 4: 13a, 13b, 16 (2nd half), 16b, 16c, 21	each names its exit; 16c carries an explicit caveat that its staging recipe is not verified here, with the four-variant fallback rather than a criterion that passes by never reaching its path

Re-measured at 7586905 (obligation: paste the output). Every count in this document, re-run in this worktree:

$ grep -rn 'describe\.buffer' builtin/ --include='*.lua' | wc -l
14
$ grep -rn bypass_intercept builtin | wc -l
21
$ grep -rn --include='*.lua' add_intercept . --exclude-dir=target | wc -l
17
$ grep -rn "is_modified()" src/ pmacs-gpu/ | wc -l
26
$ grep -rn "mark_clean()" src/ | wc -l
6
$ grep -c "dispatch_idle_for" tests/terminal_copy_mode_acceptance.rs
0

All unchanged from revision 3's readings, which were taken at ad41cf1; #189 adds no Lua and no tests. The 276 CRDT-dark figure is deliberately still not re-quoted — see §10.

---

Revision 3 — scouted against canonical githubsucks/main @ 7586905, 2026-07-28. Every claim below about pmacs was executed or read at a named line, not recalled. The reproductions in §0 and §2 are transcripts of throwaway probes run in this worktree at ad41cf1 and deleted before each commit; nothing between ad41cf1 and 7586905 touches a file they measure (#189 edits COHERENCE.md only), so they still describe the base. The counts in §1 and §2.10 are whole greps with the arithmetic shown, never | head.

Revision history

Revision 3 answers five review findings on PR #188 @ 516bd35 — three P1, two P2 — and one sweep the review asked for by class. It also withdraws one of revision 2's own decisions, reports four defects the review did not name, and says where it thinks the review is wrong.

finding	the decision
P1-1 — the proposed route cannot write to a locked buffer	Confirmed and worse than stated: the review's first repair option cannot work. run_bypass_edit calls begin_edit, whose first statement is ensure_writable() (src/buffer.rs:724-735) — but reordering begin_edit's two checks does not help, because both checks are unconditional and both return Err; the reorder only changes which error a doubly-failing call reports. Any path that admits a generated write must not call ensure_writable at all, so the design space is one option, not two. Q#GB3 is reversed: generated writes get their own run_buffer_edit arm and never touch begin_edit. The whole transaction — refusal, re-entrancy gate, unlock, write, relock, history, flag clear — becomes a single &mut Buffer method with no exit path a caller can miss (new §3.4, new Q#GB17).
P1-2 — disambiguation is not carried through name-based identity	Confirmed. The census is new §2.10: 19 units across 14 grep lines, classified by reading each enclosing function. Two sites are genuinely broken by disambiguation, and between them they have six downstream consumers, not the three the review named — listview.open's own never-capture-a-panel guard is the fourth listview consumer, and it fails silently and inverted, capturing a panel as its own q target. New Q#GB18 routes both by owned BufferId. Five further sites are the correct use of a name lookup and are named so a reviewer does not re-derive them.
P1-3 — unlock_generated is an unrestricted clear	Confirmed, and the review's second half is decisive: the binding does not achieve its stated purpose. Revision 2's Q#GB7 is withdrawn as written. The capability survives only bounded by lock provenance (new Q#GB15), moves from Stage 1 to Stage 2, and its justification narrows from "brick escape" to "the closure of the capability { generated = true } adds". Stage 1 ships no unlock because Stage 1 adds no lock capability main does not already expose.
P2-4 — Stage 1 criterion 6 passes through the side-window gate	Confirmed, and the framing cited the wrong precedent. tests/terminal_copy_mode_acceptance.rs's acc16 — which revision 2 named as the model — has no is_side assertion (grep -c dispatch_idle_for tests/terminal_copy_mode_acceptance.rs = 0; it goes through state.dispatch_idle()). The test that gets it right is tests/dired_acceptance.rs:969. The criterion is rewritten to carry both halves — the document-window premise and acc16's positive control — because dispatch_idle_for has six ways to return false and the review named one.
P2-5 — the path-backed refusal is tested only through the wrapper	Confirmed. Stage 2 criterion 6 now exercises all four surfaces — the wrapper plus each of insert / delete / replace with { generated = true } — with the misplacement as the explicit bite.

Sweep C — "a mechanism was assumed to compose with an existing guard without reading the guard" (obligation 3). Findings 1 and 3 are both that shape. Sweeping the rest of the framing for it found two more, both in the shipped primitive, and both unframed anywhere:

1. Buffer::set_generated_contents lifts a lock it did not install. It sets read_only = false unconditionally (src/buffer.rs:546), so pmacs.buffer.set_generated_contents(term_buf, "junk") on a live terminal identity buffer — whose read_only was set by src/terminal/session.rs:305 for reasons that have nothing to do with this arc — overwrites its contents and re-locks it as if the primitive owned it. This ships today. Q#GB15's provenance field closes it in the write direction as well as the unlock direction, which is what makes a second field worth its cost rather than a one-off for Q#GB7.
2. read_only is also the fold subsystem's "is this a document buffer" test. document_bytes (src/lua_bindings/fold.rs:310-318) returns None for any read_only buffer, so pmacs.fold.fold answers false with the status fold rejected: not a document buffer — and tests/folding_acceptance.rs:570-590 pins exactly that. Locking five buffer families therefore silently disables fold creation on all five. This directly contradicts Q#GB3's own stated rationale ("folding a *compilation* buffer is possible, so changing this would be a silent behaviour change to a pinned seam"): revision 2 preserved the unfold-on-edit seam while the lock it recommends kills the fold-creation seam. New Q#GB16.

Two further defects found while answering the review, neither of them of that shape and neither named by the review:

3. §1.5's correction has landed and this document was about to re-assert it. PR #189 (main @ 7586905, merged after revision 2 was pushed) corrects COHERENCE.md §14's listview consumer list to exactly the three call sites §1.5 measured, and moves the row from ✓ to ◐. §1.5, §8 and §9's "correction owed" are rewritten from owed to landed. A merged correction must be removed, not restated.
4. *help* has two independent writers, one of them Rust, and §1.4 named only the Lua one. src/help.rs:354-372 replace_help_buffer does its own find_by_name(HELP_BUFFER_NAME)-or-create, writes delete-all + insert-all through Buffer::apply_edit, and calls mark_clean() (:381) — while builtin/commands/default.lua:1229-1246 does the same thing in Lua and does not mark clean. Two owners, two copies of the name constant across the FFI boundary, one buffer. It is a fifth writer mechanism, and it is a reason Q#GB9's deferral of Class C is right rather than merely convenient.

Revision 2 answered five review findings on PR #188 @ 9101bf5 — three P1, two P2 — and two sweeps the review asked for by class rather than by item. Nothing was silently rewritten; each change is attributed below.

finding	what it changed
P1-1 — three Stage 2 criteria are non-discriminating	Stage 2 criteria 3, 4 and 5 rewritten (§6). All three passed on the pre-image: ordinary edits are already refused by the existing intercept, and Buffer::undo checks read_only before it looks at history (src/buffer.rs:1302), so "undo returns false" passes against an implementation that locks and never clears. The new wording uses a bypass write or Rust-side Buffer::is_read_only() to prove locking, and lifts the lock inside a Rust test before asserting NothingToUndo / can_undo() == false to prove clearing. Confirmed against the tree.
P1-2 — staging omits the ownership prerequisite	New §2.8 (measured), new Q#GB13, amended Q#GB5, and staging changes in §5: ownership-by-handle is now a prerequisite of the stage that locks each writer, not a follow-up. Confirmed and materially worse than the review stated — §2.8 measures that a failed pmacs.compile.run already leaves a foreign buffer permanently un-editable today, and that M-x buffer.undo is currently the only way to recover a clobbered one. This arc removes that accidental safety net, which is exactly why it cannot ship without ownership.
P1-3 — mark_clean can suppress recovery	New §2.9 and a rewritten Q#GB10. Revision 1 was wrong: it claimed is_modified "drives only the mode-line indicator and the buffer-list column". It is also read by src/autosave.rs:363 — the skip that decides whether a crash-recovery slot is written — and src/desktop.rs:302. The rule chosen and framed: a generated write refuses a buffer that has a file_path, which bounds the contents clobber and the lock as well as the flag.
P2-4 — Q#GB6 conflates byte extent with line extent	Q#GB6 rewritten. win.view_top is a line index (src/window.rs:373-374) bounded by TextView::line_count; win.cursor is a byte position bounded by Buffer::len. The clamp is now per-coordinate and ungated, matching rebuild_views_for's own shape (src/editor_core.rs:1875-1879), and §6 gains a longer-in-bytes / fewer-in-lines pin.
P2-5 — the CRDT-dark count was knowingly stale	Re-measured at ad41cf1: 276 dark (3,251 vs 3,527), with the command shown in §10. Revision 1 quoted 264, which docs/active-work.md:107-115 explicitly labels historical with "the number moves with every merge and must be re-measured, not quoted".

Sweep A — every criterion re-audited against its pre-image, not only 3–5. Two results beyond the cited items. First, eight criteria pass on main by design and their bites name a non-main pre-image; that is legitimate (docs/agent-handoff.md §5: "bite against every pre-image the fix could plausibly have taken"), but revision 1 did not say so, and an unlabelled always-green criterion is indistinguishable from a vacuous one. §6 now carries a pre-image column for every criterion. Second, Stage 1 criterion 7's stated bite was wrong: a "partial conversion" that keeps a bypass_intercept write beside the primitive does not produce a stale paint, it raises at the bypass write (§2.4, measured), so the criterion could never have failed the way it claimed. Restated as an explicit mutation bite.

Sweep B — "a capability was made public without bounding who may use it on what." Two results beyond P1-2 and P1-3. First, the pathless rule bounds what but not who: any Lua, including a third-party package, can still permanently lock *scratch* — pathless, the default buffer, and the quit target of five different code paths. Second, and decisive, the two halves of the protection are not symmetric: the intercept is removable (remove_intercept, src/lua_bindings/mod.rs:3433, used by the REPL at repl/init.lua:325-327) and the rope lock is one-way from Lua. That falsifies revision 1's stated reason for deferring the unlock ("a binding whose only consumer does not exist yet cannot be pinned"): the brick scenario is a consumer and is pinnable. Q#GB7 is upgraded from "name it, do not build it" to "ship the unlock in Stage 1."

Revision 3 keeps sweep B's observation and reverses its conclusion (review P1-3). The asymmetry is real. The brick scenario is not a recovery consumer — by the time anyone reaches for the unlock, set_generated_contents has already cleared the history — and the *scratch* exposure it turns on is one main already ships, since set_generated_contents is already public. Sweep B also stopped one step short: an unbounded clear of read_only is a strictly larger capability than the arc adds, because the same flag protects live terminal identity buffers. Q#GB7 now lands in Stage 2, bounded by Q#GB15's provenance.

Revision 1 — initial framing: the confirmed bug, the classified census, the primitive decision, staging, and acceptance.

This closes the class-wide half of the invariant Buffer::set_generated_contents opened in terminal copy mode (#178) and that docs/agent-handoff.md §4 and COHERENCE.md §14 both record as unfinished: four writer mechanisms across five buffer families still pair an erroring intercept with bypass_intercept writes over a writable rope, and every one of them is emptied by undo.

Two things the arc turns out NOT to be, both discovered by measurement:

• It is not "compile is the urgent one". compile.lua and the *search-results* panel rebind all seven undo chords to a no-op (compile.lua:219, builtin/commands/default.lua:855), so reaching them needs M-x. dired.lua and listview.lua rebind nothing, so a bare C-/ empties them. The cheap half is also the exposed half.
• It is not a Lua-only change. Two of the four mechanisms write incrementally and cannot use the shipped primitive at all, and a buffer the shipped primitive has locked refuses bypass_intercept writes (§2.4, measured) — so partial adoption is impossible and a new Rust primitive is required.
• Added in revision 2: it is not safe to lock these buffers before fixing who owns them. Three of the five writers adopt any buffer that shares their name (§2.8, measured), and the only thing that currently recovers a clobbered user buffer is M-x buffer.undo — this arc's bug. Ownership is a prerequisite, not a follow-up (Q#GB13).
• Added in revision 3: and fixing ownership is not free, because fixing it changes what buffers are called. Five sites in builtin/ recover a buffer's identity from its name; two of them break (§2.10), and one of those two — listview.open's own q-target guard — breaks inverted and silently, producing exactly the failure its comment says it prevents. Q#GB18.
• Added in revision 3: read_only is not this arc's flag. It carries three unrelated authorities (§2.11), so a capability defined over it reaches all three — which is why revision 5 removes the unlock rather than pretending identity_protected is provenance (Q#GB7), why the lock silently disables fold creation (Q#GB16), and why the shipped primitive can already overwrite a live terminal's identity buffer (Q#GB15).

---

0. The bug, reproduced

Buffer::undo (src/buffer.rs:1301) gates on ensure_writable() (src/buffer.rs:568) and nothing else. ensure_writable reads the Rust read_only field; it never consults the intercept chain. The pmacs.buffer.add_intercept(buf, function() error(name .. " is read-only") end) idiom therefore protects the edit path and leaves the history path wide open, while the owner's own bypass_intercept paint lands on the undo stack for undo to pop.

The user-reachable chain, verified end to end:

M-x buffer.undo → cmd { name = "buffer.undo" } (builtin/commands/default.lua:179) → ed.undo() → EditorCore::undo (src/editor_core.rs:2597) → Buffer::undo → ensure_writable. The chords C-/ C-_ C-4 C-x u are bound globally (builtin/keymaps/default.lua:126-136) and the menu carries it too (builtin/menus/default.lua:141). No buffer-local rebinding removes the command, which is what compile.lua's own comment already admits ("command/menu undo stays dispatchable", compile.lua:236).

0.1 Measured transcripts

Every line below is probe output from ad41cf1.

listview panel, plain C-/ through dispatch_key — no rebinding exists, so this is the whole distance from a keystroke to an empty panel:

listview BEFORE   = "H\nrow-one\nrow-two"
listview after C-/ = ""

listview panel, ordinary edit — the intercept works, which is exactly why the idiom reads as safe:

listview ORDINARY EDIT = false | intercept rejected the edit:
  builtin/runtime/listview.lua:102: *probe-panel* is read-only

dired listing, one buffer.undo:

dired BEFORE  = "/tmp/.tmpEJlp9i:\n  -rw-r--r--  1 2026-07-28 17:42 alpha.txt\n  -rw-r--r--  1 2026-07-28 17:42 beta.txt"
dired AFTER 1 = ""

*shell-command*, M-x buffer.undo through the real minibuffer (M-x, typed buffer.undo, RET):

shell BEFORE = "$ printf ...\nDirectory: ...\n\none\ntwo\n\n[shell exited with code 0]\n"
shell after M-x buffer.undo
             = "$ printf ...\nDirectory: ...\n\none\ntwo\n\n[output desynced by external edit]\n"

Read that one carefully — it is the single most important measurement in this document. The Q#CM2 revision guard noticed and appended its desync marker. It did not prevent anything: the run's exit status is gone for good, and the buffer is still non-empty. Any acceptance criterion phrased as "the buffer is not empty" passes with the bug live. See §5, Stage 2 criterion 1.

Driven programmatically to the end, the same buffer empties completely:

shell AFTER 1 undo  = "... one\ntwo\n"          (exit marker gone)
shell AFTER 13 undos = ""

*search-results*:

search BEFORE = "Searching for: fn main\n\n"
search AFTER  = ""

0.2 The fix already in the tree, and what it proves

terminal.lua is the adopter and the precedent. render_snapshot (terminal.lua:320-337) calls pmacs.buffer.set_generated_contents, and the comment at :322-336 documents this exact defect in these exact terms. claim_snapshot (:339-396) keeps the erroring intercept and set_round_trip_input, and :351-366 states the layering that this framing must preserve at every adopter:

rope-level read-only protects the daemon copy, round-trip input protects the replica copy — and neither substitutes for the other.

A locked buffer measured at ad41cf1:

bypass write after lock            = false | buffer `*probe*` (id BufferId(4)) is read-only
M-x buffer.undo after lock leaves  = "header\n"

---

1. The census, with its arithmetic

1.1 bypass_intercept — 21 grep hits, 16 write sites

grep -rn bypass_intercept builtin returns 21 lines. Five of them are prose in comments, not calls:

file:line	what it is
compile.lua:265	comment above ensure_slot's intercept
terminal.lua:304	comment in unique_snapshot_name
terminal.lua:324	comment in render_snapshot (the round-2 note)
dired.lua:478	comment above claim_handle
listview.lua:9	module header

21 − 5 = 16 actual write call sites, and the per-file arithmetic is 9 + 4 + 1 + 2 + 0 = 16:

file	writes	lines
compile.lua	9	319, 443, 454, 465, 506, 512, 642, 794, 798
builtin/commands/default.lua	4	827, 849, 1005, 1007
dired.lua	1	371
listview.lua	2	60, 61
terminal.lua	0	— (it adopted the primitive)

Two corrections to the counts this lane was briefed with. compile.lua has 9 write sites, not 10 — the tenth hit is the comment at :265. terminal.lua's two hits are both comments; it performs no bypass_intercept write at all, which is the correct state for an adopter and is worth stating because the raw grep count reads as though it still does.

1.2 add_intercept — 17 Lua sites, 6 production

grep -rn --include='*.lua' add_intercept . --exclude-dir=target returns 17 lines: 6 in builtin/, 11 under tests/fixtures/. One of the eleven (tests/fixtures/pmacs-mcp-prompts/init.lua:84) is a doc comment, so the fixture call count is 10; 6 + 10 = 16 calls across 17 lines. The six production sites:

site	buffer(s)	shape
terminal.lua:367	terminal copy snapshot	blanket read-only — adopted
dired.lua:509	every dired buffer	blanket read-only
listview.lua:101	every listview panel	blanket read-only
compile.lua:266	*compilation*, *shell-command*	blanket read-only
builtin/commands/default.lua:869	*search-results*	blanket read-only
builtin/packages/repl/init.lua:187	REPL buffers	filtering — §2.5

1.3 set_read_only — zero Lua callers, and no Lua binding

grep -rn set_read_only builtin tests returns 5 hits, all Rust test code (tests/folding_acceptance.rs:587, tests/vterm_stage1_acceptance.rs:139,175, tests/terminal_copy_mode_acceptance.rs:582,584). The stronger fact: the Lua binding table registers "add_intercept" and no "set_read_only" / "is_read_only" at all (src/lua_bindings/mod.rs:3409 is the only match in the neighbourhood). Lua cannot set read_only today. That matters for Q#GB7.

1.4 The classification, by writer mechanism

Class A — erroring intercept + bypass_intercept writes over a writable rope. This is the bug.

1. terminal.lua — copy-mode snapshot. ADOPTED (:336). Fixed.

2. dired.lua — every dired buffer. One write, in paint (:369-372): handle.buf:replace(0, handle.buf:len(), text, {bypass_intercept=true}). A whole-buffer replace already. Convertible with the shipped primitive.

3. listview.lua — *references*, *outline*, *lsp-help* (the three production listview.open callers, all in lsp.lua:2056, 2102, 2513). One writer, render (:50-62): delete-all then insert-all, which is a whole-buffer replace spelled in two ops. Convertible.

4. compile.lua — *compilation* and *shell-command*, both via ensure_slot (:258). Nine writes across five enclosing functions, and they are genuinely incremental:

   enclosing function	line	shape
   resync (:309)	319	append desync marker at end
   emit_text (:432)	443	append remainder at end
   emit_text	454	append "\n" at end
   emit_text	465	positional replace (CR overwrite)
   apply_events (:480)	506	targeted delete (erase-to-eol)
   apply_events	512	targeted delete (erase-line)
   emit_text_raw (:639)	642	append marker at end
   start_run (:746)	794	delete-all (run reset)
   start_run	798	insert header (run reset)

   NOT convertible. emit_text is a terminal emulator: it tracks slot.out_pos / slot.line_start / slot.parse_line_start as byte anchors, reads buf:slice(pos, len) between writes, and settles slot.expected_rev = buf:revision() afterwards. A whole-buffer replace destroys every one of those anchors.

5. builtin/commands/default.lua — the independent *search-results* panel (ensure_search_panel, :857). Four writes across three enclosing functions:

   enclosing function	line	shape
   search_panel_resync (:821)	827	append desync marker
   search_panel_append (:844)	849	append match batch
   pmacs.project.search (:982)	1005	delete-all (query reset)
   pmacs.project.search	1007	insert header (query reset)

   NOT convertible, for the same reason at smaller scale: the append path carries p.next_row / p.expected_rev bookkeeping.

   Do not read ensure_slot as covering this panel. It serves *compilation* and *shell-command* only; *search-results* has its own intercept, its own round-trip mark, its own resync and its own writes, and compile.lua names it only inside the is_generated_buffer predicate (:216).

Class B — filtering intercept, deliberately partly editable.

6. builtin/packages/repl/init.lua:187 — see §2.5. Shares the root cause, does not share the remedy. Out of this arc.

Class C — generated, but nothing ever claimed they were protected. Keying the inventory on bypass_intercept misses these entirely, because an unprotected buffer needs no bypass:

7. *buffer-list* — render_list (default.lua:387) writes with plain buf:delete(0, len) / buf:insert(0, body) (:403-404). No intercept, no round-trip mark. Whole-replace shape.

8. *help* — TWO independent writers, one Lua and one Rust. Corrected in revision 3; revision 2 named only the first.

   • show_help_text (default.lua:1239), plain delete-all + insert-all (:1245-1246), no intercept, no mark_clean.
   • replace_help_buffer (src/help.rs:354-383) — find_by_name( HELP_BUFFER_NAME)-or-create (:358-360), delete-all + insert-all through Buffer::apply_edit (:365, :372 — the intercept-running path, not the skip path), then Buffer::mark_clean() (:381) with the comment "The help buffer is regenerated content".

   The name constant is declared twice, independently, on either side of the FFI boundary: src/help.rs:38 (pub const HELP_BUFFER_NAME) and builtin/commands/default.lua:1226. Neither writer knows about the other; they differ on mark_clean and on which write primitive they use. Two owners for one buffer is why Class C's deferral (Q#GB9) is correct rather than merely convenient — "make *help* immutable" is not a conversion, it is first a decision about who owns it.

9. *workers* — a Rust writer, workers_buffer::render (src/workers_buffer.rs:65), using Buffer::apply_edit (not the skip-intercepts path), delete-all + insert-all, then Buffer::mark_clean() (:95). Its fan-out is a fourth mechanism: queue_generated_buffer_edits + rebuild_generated_buffer_views (src/lua_bindings/mod.rs:7142-7145).

Class C is a different defect — nothing is defeated, because nothing was claimed. It is named here so the inventory is complete and so a future reviewer does not re-derive it; §4 keeps it out of this arc.

Revision 3's correction to the inventory's headline number. This document, docs/agent-handoff.md §4 and COHERENCE.md §14 all say four writer mechanisms. Counting src/help.rs:354 — a distinct mechanism by every criterion the others are counted by (its own find-or-create, its own write primitive, its own clean-marking policy) — the honest figure across Classes A and C is five, over seven buffer families. The four-row table in the handoff is keyed on bypass_intercept and structurally cannot see it. Carried to the PR body; the handoff is not this lane's file to edit mid-flight.

1.5 The COHERENCE.md §14 correction — LANDED, not owed

Revision 3 rewrites this section from a claim into a record, because the correction merged while revision 2 was open.

Revision 1 and 2 recorded that §14's "references, outline, buffer-list, and project-search all use listview" was wrong: pmacs.listview.open has three production callers, all in lsp.lua — *references* (:2056), *outline* (:2102), *lsp-help* (:2513) — while *buffer-list* is hand-rolled in default.lua (render_list, :387) and *search-results* is the independent grep panel.

PR #189 landed exactly that correction (main @ 7586905, 0dd0bf2): §14's List bullet now names the three lsp.lua call sites, the scorecard row moves from ✓ to ◐, and the §6 picker/panel parenthetical gains *lsp-help*. Nothing is owed. This section survives only so a reader of the earlier revisions does not go looking for a correction that is already in the tree, and so the reason stays recorded: the miscount came from counting the compile.lua and dired.lua comments that cite "the listview idiom" as adoptions. Those comments are imitators, and what they imitate is the erroring-intercept-over-a-writable-rope pattern this document exists to fix.

---

2. Ground truth (measured, not recalled)

2.1 What the shipped primitive is

Buffer::set_generated_contents (src/buffer.rs:545, doc comment :507-544): lift read_only, apply_edit_skip_intercepts a single whole-buffer EditOp::Replace, clear_history() (the call is :553; :559 is the definition — a revision-2 miscitation), re-assert read_only, return the Edit. The Lua binding (src/lua_bindings/mod.rs:3079-3095) fans that Edit out via notify_buffer_edit_to_windows (:1573) after dropping the registry borrow, because the fan-out re-enters the core.

clear_history clears whichever history the buffer has: the v0.1 undo/redo stacks, and in CRDT mode CrdtState::clear_undo_history (src/crdt.rs:507), which rebinds a fresh UndoManager to the same doc because loro exposes no clear.

2.2 Why history clearing is load-bearing

The doc comment's reason is retention, not tidiness: read_only guarantees the pushed entries can never be popped, so a periodically refreshed panel accumulates full rope clones nothing will ever release. CRDT mode has the identical retention inside loro's UndoManager.

2.3 Why a bare lock is not the answer

ensure_writable guards the bypass path too (apply_edit_skip_intercepts, src/buffer.rs:1055-1056). Locking a generated buffer without giving its owner a door refuses the refresh the buffer exists for. That is why the pairing is the primitive and why there is deliberately no Lua set_read_only today.

2.4 Partial adoption is impossible — measured

This is the fact that decides the design. Once set_generated_contents has locked a buffer, a subsequent owner write through bypass_intercept is refused:

bypass write after lock = false | buffer `*probe*` (id BufferId(4)) is read-only

So compile cannot convert its run reset (start_run:794,798) to the shipped primitive and keep bypass_intercept for streaming: the first append after the reset raises. The streaming owner needs a write path that itself carries authority. Reads are unaffected — buf:slice, buf:len and buf:revision all work on a locked buffer, which is what makes an op-level solution viable at all.

2.5 The REPL: same root cause, different remedy — measured

builtin/packages/repl/init.lua:187 installs function(op) return repl._intercept(h, op) end — a filtering policy (repl._intercept, :686-726): reject edits wholly inside the history/prompt region, truncate edits that straddle the boundary, pass edits in the input region. Its own writes use a _self_write flag (with_self_write, :111-116) rather than bypass_intercept, and it has real teardown (remove_intercept, :325-327).

Does it share the bug? Yes — measured — and I am not asserting the comfortable answer:

repl BEFORE                 = "line one\nline two\n> "
repl ordinary edit at pos 0 = false | REPL: history/prompt region is read-only
                                      (insert at 0; input region begins at 20)
repl AFTER 1 buffer.undo    = "line one\nline two\n"
repl bookkeeping after undo = _history_end=18 / _prompt_end=20  (rope is 18 bytes)

Undo deleted the prompt the intercept had just refused to let anyone touch, and left _prompt_end pointing two bytes past the end of the rope. The marks (_history_end_mark, _prompt_end_mark) adjust with the rope, but _blocks[i].start_byte are plain integers maintained by hand in drop_oldest_block (:643-656) and do not.

But the remedy cannot be rope-level read_only: the input region must accept ordinary user edits, which is the whole point of a REPL. The REPL needs either an undo that consults the intercept chain, or mark-anchored blocks. Both are different work. Q#GB8: out of this arc, named deferral, with its measurement recorded above so the next lane does not have to rediscover it.

2.6 A pre-existing defect in the shipped primitive — measured

notify_buffer_edit (src/editor_core.rs:1836) updates each window's TextView and overlays. It does not clamp win.cursor or win.view_top. Only rebuild_views_for (:1865) does, and its doc comment says so explicitly (:1863-1864). set_generated_contents's binding calls the former.

cursor before = 29, len = 30
cursor after set_generated_contents(G, 'x\n') = 29, len = 2
row0 after shrink = "x"          (paint did not crash)
cursor after C-p  = 29           (motion did not recover it)

A shrinking generated write leaves the window cursor 27 bytes past the end of the buffer, indefinitely. This ships today in terminal copy mode — refresh a snapshot to a shorter one with the point low in the buffer and this is the state — and every adopter inherits it.

The two coordinates fail on different axes (review P2-4). cursor is a byte position (src/window.rs:366-367) bounded by Buffer::len(); view_top is a line index (:373-374, "First buffer line shown at the top") bounded by TextView::line_count() (src/text_view.rs:67). The transcript above is the byte case. The line case is not measured — staging it needs a scrolled window — but it is available from the types alone: a write that grows in bytes while collapsing lines invalidates view_top on a write no byte-length comparison calls a shrink. rebuild_views_for already clamps each against its own bound (src/editor_core.rs:1875-1879); the clamp added to notify_buffer_edit must do the same. Q#GB6.

2.7 What buffer.after-edit does and does not do

buf:insert / buf:delete / buf:replace do not fire buffer.after-edit; the dispatcher and daemon do (src/editor.rs:1472,2356,2500,2535, src/daemon.rs:3167). compile.lua:714 already relies on this ("hook edits don't re-fire the hook"). Consequence for §3: a generated write does not run arbitrary Lua, so the fan-out is not a re-entrancy hazard — but a scoped primitive's callback body still is, because it is arbitrary owner Lua.

2.8 Three writers adopt any buffer that shares their name — measured

The invariant already exists in this codebase; three writers simply do not honour it. terminal.lua:300-305 states it verbatim:

pmacs.buffer.create takes any caller-chosen name, so a foreign buffer may already be called *terminal-copy: sh* [...] found-by-name is NOT adoption: ownership means "this buffer is in the handle table above", exactly as in dired.

dired.lua:476-504 implements the same rule: claim_handle looks up its handle table first, and on a name collision disambiguates <2>…<99> (NAME_VARIANT_LIMIT, :474) or raises. Three writers instead adopt:

writer	line	code
listview.ensure_panel	listview.lua:95	find_buffer_by_name(name) or pmacs.buffer.create(name)
compile.ensure_slot	compile.lua:263	buffer_named(name) or pmacs.buffer.create(name)
ensure_search_panel	default.lua:861-868	name scan over pmacs.buffer.list(), then buf or create

Measured at ad41cf1, a user buffer named *references* and then a references panel:

foreign BEFORE               = "my precious notes"
foreign AFTER listview.open  = "H\nr1"
buffers named *references*ish = 1              (no disambiguation happened)
ordinary edit to MINE now    = false | intercept rejected the edit:
                               listview.lua:102: *references* is read-only

The user's buffer is clobbered and left permanently un-editable, because ensure_panel installs an erroring intercept whose handle it discards.

Compile is worse, and it is worse on a path that fails. pmacs.compile.run calls ensure_slot (compile.lua:1090) before start_run validates opts.display (:752-757). Measured:

compile.run('true', { display = 'bogus' })
  = false | compile.lua:754: compile.run: unknown display "bogus"
foreign *compilation* contents after the FAILED call = "my precious notes"
ordinary edit to MINE after the FAILED call
  = false | intercept rejected the edit: compile.lua:267: *compilation* is read-only

A call that raised and did nothing else left the user's buffer uneditable. Q#GB5's revision-1 recommendation — an empty set_generated_contents at the end of ensure_slot — would make that same failing call empty the buffer and lock the rope. Q#GB5 is amended accordingly.

Why this is a prerequisite and not a follow-up. Today the clobber is recoverable, and the thing that recovers it is this arc's bug:

after clobber = "H\nr1"
after undo 1  = ""
after undo 2  = "my precious notes"

M-x buffer.undo is currently the only way back. After adoption the rope is read_only, the history is cleared by the same call that wrote, and §1.3 measured that no Lua binding can clear read_only. The arc therefore converts a recoverable clobber into an unrecoverable one, and it removes the accidental safety net in the same commit that removes the need for it. Q#GB13.

Dired needs none of this work — it already disambiguates — which is why it is the cheaper of Stage 1's two adopters despite being the newer one.

2.9 is_modified reaches autosave and desktop persistence — a revision-1 error

Revision 1 stated that the flag "drives only the mode-line indicator and the buffer-list column". That is wrong, and it was wrong because the sweep was grep -rn '\.modified' builtin plus a narrow src path rather than grep -rn 'is_modified' src. The full sweep finds two more consumers, both load-bearing:

• src/autosave.rs:359-364 — the per-buffer skip: let Some(path) = buf.file_path() else { continue }; then if !buf.is_modified() { continue; }. A clean buffer gets no crash-recovery slot written.
• src/desktop.rs:298-303 — SavedBuffer { path, modified: b.is_modified() }, again only for buffers with a file_path.

Both gate on file_path() being Some before they read the flag. That is the fact Q#GB10's revised rule turns on.

Revision 3: the sweep was under-run a second time, and the arithmetic is stated here so it is not under-run a third. grep -rn 'is_modified()' src/ pmacs-gpu/ returns 26 lines. Removing the accessor definition and the 18 src/buffer.rs unit-test assertions leaves 7 production consumers:

consumer	load-bearing?
src/autosave.rs:363	yes — the crash-recovery skip (§2.9 above)
src/desktop.rs:302	yes — the persisted SavedBuffer.modified
src/editor.rs:4129	no — the TUI mode-line *
src/semantic_render.rs:1716	no — the semantic frontend's statusline payload
src/help.rs:131	no — the Modified: line of describe-buffer text
src/instance_buffer.rs:401	no — an assertion, not a read
src/lua_bindings/mod.rs:1262, :6359	no — buf:is_modified() and describe.buffer().modified, which default.lua:395 renders

Revision 1 said two consumers; revision 2 said four; the true figure is seven, of which two are load-bearing. Both new ones (semantic_render.rs:1716, help.rs:131) are display, so Q#GB10's conclusion is unchanged — but the conclusion was reached twice from an incomplete count, and only the arithmetic makes that visible.

Also found in the same sweep, and reused below: mark_clean() has six callers (grep -rn 'mark_clean()' src/, minus the definition). src/instance_buffer.rs:95,114, src/workers_buffer.rs:76,95 and src/help.rs:381 are all generated-buffer writers that already mark clean; src/editor_core.rs:1967 is the save path. So the convention Q#GB10 adopts is established by three Rust writers, not the one revision 1 cited.

2.10 The name-based identity census, with its arithmetic

Review P1-2 named two consumers. This is the whole set, and the classification comes from reading each enclosing function, never from the grep line.

Scope. A site is in scope when a disambiguated name would change its answer: it either (i) recovers a buffer's identity by comparing that buffer's name against an expected value, or (ii) keys a table by a buffer name. pmacs.describe.buffer(id).name is the only Lua surface that yields a buffer's name — buffer_info_table (src/lua_bindings/mod.rs:6352-6364) sets name, length, modified, view_count and nothing else, and no other binding exposes it — so grep -rn 'describe\.buffer' builtin/ --include='*.lua' is a complete frontier for builtin/.

The arithmetic. That grep returns 14 lines. Five of them are the bodies of shared helpers rather than decisions:

helper	file:line	callers
buffer_name	terminal.lua:276	1 (:293)
buffer_named	terminal.lua:283	2 (:308, :311)
buffer_named	dired.lua:196	3 (:491, :495, :914)
buffer_named	compile.lua:194	2 (:263, :1052)
find_buffer_by_name	listview.lua:32	2 (:95, :190)

14 − 5 helper bodies = 9 direct sites; the five helpers expand to 1 + 2 + 3 + 2 + 2 = 10 call sites; 9 + 10 = 19 units, and the five classes below partition them 2 + 3 + 2 + 5 + 7 = 19.

Class 1 — BROKEN by disambiguation. In scope for this arc (Q#GB18).

1. listview.lua:42-44, panel_for_current_buffer. panels is written panels[name] = p at :97 with the requested name and read return panels[d.name] at :44 with the actual name. A panel created as *references*<2> can never resolve its own record. Four consumers, not the three the review named:

   consumer	line	what breaks
   listview.visit (RET/SPC)	:150	returns early; RET does nothing
   listview.refresh (g)	:161	returns early; g does nothing
   listview.quit (q)	:177	returns early; q does nothing
   listview.open's capture guard	:118-123	fails inverted, and silently

   The fourth is the one worth reading twice. listview.open captures the return target with if active and not panel_for_current_buffer() then p.prev = active end, and the comment above it states the intent: "never another panel (chained panels would trap q in a loop; restore targets the last real buffer)". When panel_for_current_buffer() cannot recognise a disambiguated panel it returns nil, the guard reads as "the current buffer is not a panel", and the panel is captured as its own q target — exactly the loop the guard exists to prevent, produced by the guard. The other three fail closed and visibly; this one fails open and quietly, which is why it needs its own criterion rather than riding on the other three.

2. compile.lua:214-216, pmacs.compile.is_generated_buffer. return d.name == COMPILATION or d.name == SHELL_OUT or d.name == SEARCH_RESULTS. Two consumers, both the Q#CM11 never-capture-a-generated-buffer q-target discipline: compile.lua:762 (compile's own capture) and default.lua:993-994 (the search panel's). A disambiguated *compilation*<2> is not recognised, so it gets captured as a q target and q returns the user to a generated buffer.

   Note that compile.lua:232's slot.name == COMPILATION is not in this class and needs no change: slot.name is the record's own field, set from the module constant ensure_slot was called with, and never from a buffer's actual name.

   Note also that slots (compile.lua:185) is not broken by disambiguation, unlike panels: it is keyed by the module constant at both write (:262) and read (:259, :923, :1104), and the only buffer→slot direction is slot_for_buffer (:200-206), which compares slot.buf == buf by id. Compile got this half right and listview did not; the census is what makes that visible, and it means Q#GB18's compile work is one predicate, not a table rewrite.

Class 2 — the find-by-name adoption Q#GB13 already removes. In scope, already framed. listview.lua:95 (Stage 1), compile.lua:263 (Stage 2), default.lua:863 (Stage 2). 3 sites.

Class 3 — the same adoption defect in Class C families. Out of arc (Q#GB9), named so it is not re-derived: default.lua:380 (find_list_buffer, *buffer-list*) and default.lua:1231 (find_or_create_help_buffer, *help*). 2 sites. The Rust-side instance of the same shape, src/help.rs:358-360, is outside this census's builtin/ frontier and is recorded in §1.4 instead.

Class 4 — collision probes: the CORRECT use of a name lookup. dired.lua:491, :495 (claim_handle's <2>…<99> walk); terminal.lua:308, :311 (unique_snapshot_name); terminal.lua:293 (snapshot_base_name, which derives a new name from a name and recovers no identity). 5 sites. These are the shape Q#GB13 asks the other three writers to adopt, so they are the reference implementation, not debt.

Class 5 — correct by construction; a disambiguated name does not change the answer. 7 sites.

• default.lua:391 — render_list prints d.name in a column. Display.
• default.lua:604 — switch-to-buffer matches the name the user typed, sourced from the same registry that would show a disambiguated name. Correct precisely because it is name-based.
• The five *scratch* fallbacks — default.lua:581, :1145, listview.lua:190, compile.lua:1052, dired.lua:914. *scratch* is an unowned shared buffer that no writer in this arc disambiguates, so its name is its identity.

One flag on Class 5, carried rather than fixed. Those five *scratch* fallbacks are five independent copies of one find-or-create, and they are correct only while *scratch* stays unowned and undisambiguated. If a future lane gives *scratch* an owner — a plausible move, since it is the quit target of all five paths — all five break together and nothing in the tree connects them. Named in §8.

2.11 read_only is one boolean serving three policies

Revision 2 treated read_only as this arc's flag. It is not, and both review P1-3 and sweep C turn on that.

policy	who sets or reads it	what it means there
generated lock	Buffer::set_generated_contents (src/buffer.rs:546, :554)	"the owner's write path is the only writer"
terminal identity	src/terminal/session.rs:305, at TerminalSession::open	"the host may not edit this at all — not by edit, not by undo, not by remote CRDT import"
"is this a document buffer?"	document_bytes (src/lua_bindings/fold.rs:310-318), reading it	Q#FD11's foldability test

Buffer::set_read_only's own doc comment (src/buffer.rs:496-502) describes only the second, and the third is a reader that was written when the second was the only writer — tests/folding_acceptance.rs:570-573 says so in as many words: "terminals are read-only, so a read-only buffer is not foldable."

Six paths gate on the flag through ensure_writable (src/buffer.rs:568): begin_edit (:725), apply_edit (:773), apply_remote_crdt_op (:845), apply_edit_skip_intercepts (:1056), undo (:1302), redo (:1410). That breadth is the point of the flag and is not in question. What is in question is that a single boolean carries three unrelated authorities, so a capability defined over it — in either direction — necessarily reaches all three. Q#GB15 and Q#GB16 are the two consequences.

---

3. The primitive decision (Q#GB1)

The question this arc exists to answer: what write primitive do compile.lua and the search panel need?

3.1 Recommendation

Buffer::apply_generated_edit(op: EditOp) -> GeneratedOutcome — one authorized op at a time — exposed to Lua as a new option key on the mutators that already exist:

buf:insert(pos, text,   { generated = true })
buf:delete(start, end_, { generated = true })
buf:replace(s, e, text, { generated = true })

Semantics, per call, entirely inside one with_registry_mut and one &mut Buffer method — the exact ordering, including every error path, is §3.4, which revision 3 adds because review P1-1 showed revision 2 had no workable one. The binding then fans every outcome that carries an Edit out through notify_buffer_edit_to_windows after the borrow has dropped: Applied and NoOp before returning success, and AppliedThenFailed before returning its error. Revision 5 makes that third arm explicit because collapsing the outcome to Result inside Buffer discarded the only value capable of updating window caches and replica mirrors.

Buffer::set_generated_contents(bytes) is reimplemented as apply_generated_edit(Replace { range: 0..len, bytes }). Its Lua name and call signature stay fixed; its Rust return changes from Result<Edit, BufferError> to GeneratedOutcome, and its direct Rust tests match on the outcome. The change is required: a wrapper that collapses AppliedThenFailed to Err before the binding sees it loses the applied Edit and cannot satisfy the fan-out contract. It becomes the whole-buffer spelling of one primitive rather than a second primitive.

One sentence for why it wins: it is the only candidate in which the buffer is never observably unlocked, because the lift and the re-assert happen inside a single registry borrow with no Lua in between — so there is no flag to clear on an error path, no yield to defend against, and nothing for a reviewer to audit site by site.

3.2 Why the alternatives lose

A. append_generated_contents — provably insufficient. compile.lua does a positional replace at emit_text:465 (the CR overwrite that makes progress bars work) and two targeted deletes at apply_events:506,512 (erase-to-eol, erase-line). Append cannot express any of the three. Dead on the census.

B. Scoped with_generated_writes(buf, fn) — the pattern this project has already been burned by. It is the cheapest on history (one clear_history per scope instead of one per op) and that is its only real advantage. Against it:

• The unlocked interval is the callback's whole duration, and the callback is arbitrary owner Lua. Drawn loosely around start_run (compile.lua:746), the scope spans pmacs.window.switch_buffer(buf) and pcall(pmacs.process.spawn, spec) — the buffer would be writable across a process spawn. Drawn tightly, compile needs four separate scopes (start_run's reset, resync, feed_bytes, finish_run), each needing its own audit for what the body reaches.
• Correctness reduces to "a flag cleared on every exit." That is the exact shape docs/agent-handoff.md §5 and #155 record as a repeat offender, and the REPL already had to defend its own version of it: with_self_write (repl/init.lua:111-116) wraps in pcall specifically because "a single failed write would leave the bypass on for every subsequent user edit". Adding a second instance of a pattern the tree already documents as fragile is a poor trade for one saved clear_history per batch.
• Yield is an error here, which helps but does not rescue it. A Lua callback that yields across the Rust boundary raises attempt to yield across C-call boundary (observed in this worktree while probing pmacs.dired.open), so a yielding body surfaces as Err — but the relock must still run on that path, which is the same obligation.
• It is strictly harder to review: a per-site scope audit versus a mechanical option-key change at 16 call sites.

C. Standalone generated_edit(buf, op). Identical semantics to the recommendation, worse ergonomics: it re-implements the three-op argument parsing that buf:insert/delete/replace already own, and turns adoption from an option-key change into a rewrite of 16 call sites. Recommended only if the user objects to { generated = true } sitting beside { bypass_intercept = true } in the same options table.

D. Make Buffer::undo/redo consult the intercept chain. This would fix all six families at once, including the REPL, and it deserves an explicit rejection rather than silence. Against it: (i) there is no EditOp to hand the chain — v0.1 undo is a whole-rope swap (src/buffer.rs:1327-1328) and CRDT undo is materialize-and-replace (undo_crdt_mode, :1365), so the chain would have to be given a synthetic op it was never designed to see; (ii) it changes behaviour for every intercept in the tree, including the transforming ones (auto-pair, lean-input, the REPL's truncation) which have no business rewriting an undo; (iii) an erroring intercept becomes a new Lua-raise failure path out of EditorCore::undo, which today cannot fail that way. It may still be the right answer for the REPL specifically — recorded in Q#GB8's deferral, not adopted here.

3.3 The four questions the recommendation must answer

How many Edits are fanned out, and when? One per generated op that reaches an Edit, immediately, by the binding. That includes AppliedThenFailed: the rope and CRDT have already changed, so the binding fans its carried Edit out and only then returns the carried error. compile.lua's emit_text fast path emits one insert for a whole output batch, so a typical feed_bytes produces one to three ops; a CR-heavy progress bar produces more. Successful calls retain today's fan-out cardinality; the new failure arm adds the notification that today's early Err loses, because that error can follow a real mutation.

Per-op or per-scope history clearing? Per op, and it is cheap by construction: because read_only is re-asserted immediately, at most one v0.1 undo entry can exist when the clear runs and the redo stack is always empty, so the v0.1 clear is O(1). In CRDT mode the clear rebinds a fresh UndoManager (CrdtState::clear_undo_history), and create_undo_manager (src/crdt.rs:154) is UndoManager::new(doc) plus set_max_undo_steps — a subscription registration, not a document copy, so it is O(1) in document size too. Measurement obligation, not a claim: Stage 2 must show a streaming compile run does not regress against the existing compile-mode timings in both configurations. If it does, the escape hatch is to suppress recording rather than clear it — recorded as a named deferral rather than designed speculatively.

CRDT-mode behaviour? Identical to set_generated_contents today on success, and corrected on the post-apply error path. The Edit carries crdt_op when the buffer is CRDT-backed, and notify_buffer_edit_to_windows queues it via queue_daemon_origin_crdt_op (src/lua_bindings/mod.rs:1582) so replica mirrors import the owner's write. AppliedThenFailed must carry that same Edit; returning its error before notification would leave the authoritative CRDT changed while every replica missed the op. History clearing goes to loro's UndoManager.

How do the returned edits reach the fan-out without a live registry borrow? By construction, extending #178's shape: run_bypass_edit (src/lua_bindings/mod.rs:1445) closes its with_registry_mut before returning, and the mutator bindings call notify_buffer_edit_to_windows afterwards. run_generated_edit (§3.4) occupies the same position and closes its borrow the same way, but returns the whole GeneratedOutcome. A binding-level finisher handles it:

match run_generated_edit(lua, id, op) {
    Applied(edit) | NoOp(edit) => {
        notify_buffer_edit_to_windows(lua, id, &edit);
        Ok(edit)
    }
    AppliedThenFailed { edit, error } => {
        notify_buffer_edit_to_windows(lua, id, &edit);
        Err(error.into())
    }
    Rejected(error) => Err(error.into()),
    Diverged(error) => {
        surface_divergence(lua, &error);
        Err(error.into())
    }
}

The helper is shared by the three mutators and the set_generated_contents binding so one of the four cannot forget the failure fan-out.

3.4 The transaction, and why revision 2 had none (review P1-1)

Revision 3 adds this section. It is the substance of P1-1 and it reverses Q#GB3.

The defect, confirmed at line level. Revision 2's Q#GB3 routed generated writes "through run_buffer_edit's bypass arm". That arm is run_bypass_edit (src/lua_bindings/mod.rs:1445-1454), whose first act on the buffer is buf.begin_edit(), and begin_edit's first statement is self.ensure_writable()? (src/buffer.rs:724-725). A generated write must pass while read_only is set — that is the entire point — so every generated write after the first would be refused, and for compile, whose Q#GB5 lock is installed during ensure_slot, even the first streaming write would be refused. As written, revision 2's design was dead on arrival at every buffer it governs.

The review offers two repair options; one of them cannot work, and saying so is the first design decision. The suggestion to reorder begin_edit's two checks does not repair anything:

pub fn begin_edit(&mut self) -> Result<(), BufferError> {
    self.ensure_writable()?;                       // src/buffer.rs:725
    if self.editing_in_progress { return Err(ConcurrentEdit { .. }); }
    self.editing_in_progress = true;
    Ok(())
}

Both checks are unconditional and both return Err. Reordering changes only which error a call that fails both reports; a locked buffer is still refused, one line later. It is also not free: at least one shipped test asserts on the text of that error (tests/dired_acceptance.rs:999, status(&s).contains("read-only")), and BufferError::ReadOnly and ConcurrentEdit render differently (src/buffer.rs:1794, :1824-1831). Any path that admits a generated write must not reach ensure_writable at all. So there is one option, not two: a separate entry point. Recorded as a disagreement with the review rather than complied with silently.

Where the concurrency gate lives. Inside Buffer, in the generated path itself, duplicating begin_edit's second check and not its first. It is not exposed as a public begin_generated_edit: making it a public pair would recreate at the binding layer the exact "a flag cleared on every exit" shape §3.2 rejects candidate B for. One method, one exit set, nothing for a caller to forget.

What that does to the ConcurrentEdit contract. Nothing observable, and the contract is still needed. Two directions:

• Inward (something re-enters during a generated write): impossible by construction, and this is worth stating because it is what makes the transaction safe to hold across a single borrow. apply_edit_skip_intercepts runs View::on_edit, never View::intercept_edit (src/buffer.rs:1055-1060), and LuaInterceptView (src/lua_bindings/mod.rs:1755-1798) implements only intercept_edit — it inherits View::on_edit's default no-op body (src/view.rs:252-254). Of the seven production on_edit implementations (grep -rn 'fn on_edit' src/: text_view.rs:150, fold.rs:274, overlay.rs:248, syntax.rs:1637, plus test doubles) none calls into Lua. A generated write runs no Lua, so nothing can re-enter it.
• Outward (a generated write issued from inside a managed edit on the same buffer): entirely possible — a Lua intercept body on buffer X calling X:insert(pos, s, { generated = true }) — and it must still fail. run_managed_edit phase 2 runs that body with the registry borrow released and its InterceptContext already snapshotted (src/lua_bindings/mod.rs:1477-1487); a generated write landing in between would leave phase 3 applying an op computed against a rope that no longer exists. So the generated path sets and clears editing_in_progress exactly as begin_edit/end_edit do, and this case surfaces BufferError::ConcurrentEdit unchanged.

Cleanup is driven by an explicit outcome, NOT by inferring one from revision. Rewritten in revision 4 (review round 3, P1-1); revision 3's predicate was wrong in three directions and this section says so before it says anything else.

Revision 3 wrote if self.revision() != rev_before { clear_history(); … } and argued the predicate was exact because revision bumps after the undo push and before the on_edit broadcast. That argument is sound for the v0.1 undo stack and does not extend past it. Three failures:

1. A successful no-op keeps history that the contract says cannot exist. An empty buffer can carry substantial undo history — insert, then delete back to empty. set_generated_contents("") then takes the no-op arm (src/buffer.rs:1245-1253), returns Ok, never bumps revision, so revision 3 skipped both the clear and mark_clean: the buffer ends locked, modified, and carrying poppable history that a later Rust-side lift can re-expose. Shipped set_generated_contents clears unconditionally (:551-553), so revision 3's predicate was a regression of an existing contract, not a refinement of one. And it lands on a path this document prescribes: Q#GB5's ensure_slot lock is exactly set_generated_contents(slot.buf, "").
2. CRDT mutation happens upstream of revision entirely. apply_to_crdt_then_normalize_bytes runs before the rope edit (src/buffer.rs:1194-1201), and for EditOp::Replace it is two ops — crdt.delete, then crdt.insert (:1140-1163). The code's own comment states the hazard: if the delete succeeds and the insert fails, "the CRDT is mid-transaction (range deleted but replacement not inserted) and the rope is unchanged. This is an invariant violation." In that state revision has not advanced, so revision 3's predicate neither clears the CRDT's history nor notices the rope ≡ CRDT projection divergence — in precisely the case that most needs both.
3. Found while fixing the other two: revision 3's unconditional relock locks a buffer nobody wrote to. b:insert(mid_codepoint, "x", { generated = true }) on a fresh, writable *scratch* is rejected by the CRDT after the unlock, and revision 3 then set read_only = true on the way out. The caller gets an error and a locked buffer it never successfully wrote. Not reported by the review; same class as the other two (a rule derived for one case applied to a case whose ordering differs).

The outcome value. run_rope_edit_and_broadcast reports what it actually did instead of leaving its caller to guess. This is a real signature change and not a free one — see the cost note below.

/// What a generated write did, reported by the apply rather than
/// inferred by its caller. The variant, not `revision`, selects cleanup.
pub enum GeneratedOutcome {
    /// The rope changed and an undo entry exists.
    Applied(Edit),
    /// A semantic no-op: rope unchanged, no undo entry pushed. The call
    /// SUCCEEDED, so the buffer is now a generated buffer.
    NoOp(Edit),
    /// Failed with rope, CRDT and history exactly as they were. This also
    /// includes a deliberate no-op whose `on_edit` notification failed:
    /// view side effects may have run, but no buffer/CRDT mutation exists
    /// to fan out.
    Rejected(BufferError),
    /// The rope changed and a later stage failed. The Edit must survive so
    /// windows and replica mirrors observe the mutation before the error.
    AppliedThenFailed { edit: Edit, error: BufferError },
    /// CRDT only: the CRDT was mutated and the rope was not.
    /// `rope ≡ CRDT projection` no longer holds. See the staging rule
    /// below — this is NOT "delete succeeded, insert failed".
    Diverged(BufferError),
}

Revision 6 (round 5, P1-2): Diverged is selected by whether the CRDT was MUTATED, not by which of two ops failed. Revision 5 specified a "delete→insert classifier", which recognises exactly one failure point. The CRDT routing has three, and the third is shared by every op shape: export_updates_since runs after every successful CRDT mutation and before the rope is touched (src/buffer.rs:1173), and it returns Result. So:

op	failure point	CRDT state	revision 5 said	correct
Replace	crdt.delete fails	untouched	Rejected	Rejected
Replace	crdt.insert fails after delete	mutated	Diverged	Diverged
Replace	export_updates_since fails	mutated	Rejected ✗	Diverged
Insert	crdt.insert fails	untouched	Rejected	Rejected
Insert	export_updates_since fails	mutated	Rejected ✗	Diverged
Delete	crdt.delete fails	untouched	Rejected	Rejected
Delete	export_updates_since fails	mutated	Rejected ✗	Diverged

Revision 5 got three of seven rows wrong, and each wrong row applies Rejected's cleanup — which restores a fresh buffer to writable while the CRDT and rope disagree. That is the same harm P1-3 of round 4 withdrew the four-variant fallback to prevent, reintroduced through a classifier that was too specific.

The rule, stated so it cannot miss a future failure point. The routing function tracks one bool:

let mut crdt_mutated = false;
// ... immediately after EACH successful crdt.insert / crdt.delete:
crdt_mutated = true;
// ... and every `?` becomes an explicit map that carries the flag out:
//     Err(e) => return Err((e, crdt_mutated)),

apply_to_crdt_then_normalize_bytes returns Result<CrdtRoutingResult, (BufferError, bool)>, and the caller maps (e, false) → Rejected(e) and (e, true) → Diverged(e). The classifier is therefore total by construction: any failure added later anywhere downstream of a successful CRDT op is classified Diverged without anyone remembering to extend a list of op shapes. That property is the point — revision 5's classifier was correct for the cases it enumerated and wrong because enumeration was the wrong mechanism.

run_rope_edit_and_broadcast returns this richer outcome. apply_edit and apply_edit_skip_intercepts map it back to their existing Result<Edit, BufferError> API, preserving ordinary callers' surface; the generated path retains it through cleanup and the borrow-free binding finisher. The enum is public because the existing public set_generated_contents method and the new apply_generated_edit method return it; their Rust docs state that a higher layer that owns window or replica state must fan out every edit-bearing variant before handling its success/error.

That mapping deliberately does not claim to repair the same post-apply notification loss for every ordinary Rust edit API. Their public Result<Edit, BufferError> surface cannot carry both the edit and the later view error, and changing all of those callers is broader than generated-buffer immutability. The pre-existing ordinary path is named in §8 rather than hidden by the helper refactor.

AppliedThenFailed must also finish the BUFFER-attached broadcast, not only the window and replica fan-out. New in revision 6 (round 5, P1-4). Revision 5 solved the fan-out that happens after the registry borrow drops — windows and replica mirrors — and left the loop that happens inside it broken. on_edit broadcasting stops at the first error:

// src/buffer.rs:1286-1288
for (_, view) in views.iter_mut() {
    view.on_edit(self, &edit)?;          // <- returns on the FIRST Err
}

Buffer-attached views run in attach order, and the erroring one is rarely last: SyntaxHighlightView, ParseView, FoldStoreTranslator, BufferStyleSpanTranslator and DiagnosticView all maintain byte offsets in on_edit. Any of them sitting after the failing view keeps pre-edit offsets over a post-edit rope — the same stale-index class docs/agent-handoff.md §4 records for the window TextView, one layer down. Revision 5's fan-out cannot reach them; they are not windows.

The decision: continue the broadcast and retain the first error.

let mut first_err = None;
for (_, view) in views.iter_mut() {
    if let Err(e) = view.on_edit(self, &edit)
        && first_err.is_none()
    {
        first_err = Some(e);
    }
}

FOUR broadcast sites, not one. The review named one; sweep F found the other three, and every one of them is reached by this arc. Measured:

$ grep -n "view.on_edit(self, &edit)?" src/buffer.rs
1033:            view.on_edit(self, &edit)?;
1250:                view.on_edit(self, &edit)?;
1288:            view.on_edit(self, &edit)?;
$ grep -n "view.on_edit(self, edit)?" src/buffer.rs
1543:                view.on_edit(self, edit)?;

site	enclosing fn	who reaches it
:1288	run_rope_edit_and_broadcast	apply_edit, apply_edit_skip_intercepts — the one the review named
:1250	run_rope_edit_and_broadcast, the no-op early-return arm	every NoOp outcome — including Q#GB5's ensure_slot lock, which is exactly set_generated_contents(buf, "")
:1033	run_remote_rope_stages	apply_remote_crdt_op — a replica importing the owner's write
:1543	broadcast_on_edit	undo and redo — the command this whole arc exists because of

Two of the three the review did not name are load-bearing for this arc specifically. :1250 is the path NoOp takes, and §3.4 treats NoOp as a success that leaves the buffer generated-locked — but a view erroring there today both strands the later views and converts the success into an Err. :1033 is how a locked generated buffer's write reaches a replica at all, so a strand there is a mirror that paints from stale offsets on the frontend the round-trip mark exists to protect.

All four change, to the same continue-and-retain-first-error shape. A fix applied to one, or to two, would leave the same rope with different broadcast semantics depending on which entry point wrote to it — the defect class that has now produced five findings across four rounds.

Why this changes the shipped path for ordinary edits too, deliberately. The alternative is a generated-only broadcast variant, which would give one rope two broadcast semantics depending on who wrote to it. That is the exact defect class that has now cost this arc four review findings across three rounds — a rule derived for one mechanism and applied to a second whose ordering differs. Continuing is also strictly better on its own terms: no caller benefits from a view being skipped, and a view that errors already cannot veto the rope mutation, which happened three stages earlier. Blast radius, named: apply_edit, apply_edit_skip_intercepts, undo, redo. The observable change is that a later view now sees an edit it previously missed; no error is swallowed, because the first is retained and returned.

The cleanup each variant triggers. entry is the read_only value observed on entry.

outcome	history	mark_clean	read_only after	binding action
Applied	cleared	yes	true	fan out Edit, then Ok
NoOp	cleared	yes	true	fan out Edit, then Ok
Rejected	untouched	no	restored to entry	Err, no fan-out
AppliedThenFailed	cleared	no	true	fan out carried Edit, then Err
Diverged	untouched	no	true	distinct surfaced Err, no rope Edit

editing_in_progress is cleared on all five, unconditionally.

Why NoOp clears — the rule that replaces revision 3's. The invariant is a property of the buffer, not of the write: a buffer that is generated-locked carries no history. Every outcome that leaves the buffer generated-locked therefore clears, and NoOp leaves it locked because the call succeeded. Rejected is the only success-shaped exception and it is not one — the buffer is not newly generated, so there is nothing for the invariant to apply to. Stated this way the rule needs no reference to revision at all, which is what makes it mode-independent.

Why Rejected restores rather than relocks. Revision 3's "relock unconditionally" was justified by "the contract is leave it genuinely immutable" — true of a write that happened. A refusal is not a write. Restoring the entry value keeps the refusal total: a fresh buffer stays writable, an already-generated buffer stays locked, and no caller can lock a buffer by failing to write to it.

Why Diverged must surface rather than be cleaned. This is the variant the review specifically asks about, and the honest answer is that this arc cannot fix it:

• The rope is intact and the CRDT is not. Nothing local reconstructs the deleted range — loro exposes no rollback at this seam, and the export_updates_since that would name the delta runs after both ops.
• The rope still contains the pre-edit bytes, but the CRDT UndoManager is the only existing CRDT-native inverse record for the successful delete. Clearing it would discard the one operation a later repair lane might use to reconcile the document. So Diverged clears nothing; this arc does not invoke that undo automatically.
• Locking stops further local edits compounding the divergence. So read_only = true. But see immediately below: revision 5 called this "the strongest available containment" and that was wrong.
• It returns a distinct BufferError variant rather than reusing CrdtRejected, and the Lua binding surfaces it via pmacs.editor.set_status rather than swallowing it. A caller must be able to tell "your op was refused, nothing happened" from "this buffer's CRDT and rope no longer agree."

read_only does NOT contain a divergent CRDT, and revision 5 asserted that it did. Corrected in revision 6 (round 5, P1-3). The outbound snapshot paths do not consult read_only at all — they read crdt_state() and export directly:

• initial attach, src/daemon.rs:2563-2578: buf.crdt_state() then crdt.export_snapshot(), sent as InstanceMessage::BufferSnapshot;
• buffer-follow, export_buffer_snapshot at src/daemon.rs:2693-2708: registry.get(buffer_id) → buf.crdt_state()? → crdt.export_snapshot().

Neither reads read_only, and neither can — read_only means "no local edits", which is a statement about inbound mutation. A replica attaching after the divergence therefore receives the divergent CRDT as its authoritative document, while every daemon window still paints the old rope. The lock contains exactly the direction that was already safe and none of the direction that was not.

The fix: a quarantine flag consulted by snapshot export. Buffer gains one more private field, set only by the Diverged cleanup arm and never cleared by this arc:

/// Set when a CRDT mutation landed without its rope counterpart.
/// While true this buffer's CRDT must not be published: it no longer
/// projects to what any window shows.
crdt_quarantined: bool,

Three consumers, all of which already have a skip path so the change is a guard rather than a new control flow:

1. src/daemon.rs:2563-2578 — the initial-attach loop already continues on crdt_state() returning None and on export failure; quarantine takes the same continue, with its own log line.
2. export_buffer_snapshot (src/daemon.rs:2693-2708) already returns Option; quarantine returns None, with its own log line.
3. queue_daemon_origin_crdt_op (src/lua_bindings/mod.rs) must not queue ops from a quarantined buffer, for the same reason — a delta on top of a divergent document propagates the divergence rather than the edit.

Why quarantine rather than immediate repair. Repair means either invoking the CRDT's own undo of the landed op or re-deriving the document from the rope, and both are decisions about which side wins that this arc has no standing to make: the rope is what the user is looking at, the CRDT is what replicas already hold, and if any replica has already received the divergent snapshot the answer changes again. Quarantine is the minimal correct action — it stops the divergence spreading and leaves every input a repair lane would need intact. Repair stays deferred (§8), and revision 6 does not claim otherwise.

What quarantine costs, stated: a buffer that hits Diverged stops collaborating. Replicas attaching later see no snapshot for it and continue to show whatever they had. That is a visible degradation, and it is the correct one — the alternative is silent disagreement between what the user sees and what a replica edits.

Scope, stated plainly: Diverged is a PRE-EXISTING hazard this arc exposes, not one it creates. apply_edit and apply_edit_skip_intercepts reach the same two-op replace today and report it as an ordinary CrdtRejected; nothing in the tree distinguishes it. Making it a named variant is this arc's contribution; repairing the divergence is not, and is recorded in §8 as its own lane. Splitting a CRDT Replace into a single transactional op, or reconciling the two, is loro-level work with no bearing on generated buffers specifically.

The ordering, with every exit path named. The method returns the outcome intact; only the borrow-free binding finisher above converts it to Lua success/error after performing any required fan-out.

pub fn apply_generated_edit(&mut self, op: EditOp<'_>) -> GeneratedOutcome {
    // (1) Q#GB10: path-backed refusal. Before any state change.
    if self.file_path.is_some() {
        return Rejected(GeneratedWriteOnFileBuffer { .. });
    }
    // (2) Q#GB15: this buffer's read_only is an identity protection.
    if self.identity_protected { return Rejected(ReadOnly { .. }); }
    // (3) re-entrancy gate — begin_edit's SECOND check, not its first.
    if self.editing_in_progress { return Rejected(ConcurrentEdit { .. }); }
    // (4) bounds pre-validation, so an invalid range costs nothing.
    if let Err(error) = self.validate_op_bounds(&op) {
        return Rejected(error);
    }

    let entry_read_only = self.read_only;
    self.editing_in_progress = true;
    self.read_only = false;                       // the ONLY unlocked interval
    let outcome = self.apply_generated_inner(op); // -> GeneratedOutcome
    match &outcome {                              // (5) per-variant cleanup
        Applied(_) | NoOp(_) => {
            self.read_only = true;
            self.clear_history();
            self.mark_clean();
        }
        AppliedThenFailed { .. } => {
            self.read_only = true;
            self.clear_history();
        }
        Diverged(_)          => { self.read_only = true; }
        Rejected(_)          => { self.read_only = entry_read_only; }
    }
    self.editing_in_progress = false;             // (6) unconditional, all paths
    outcome                                         // binding owns fan-out + conversion
}

exit	when	read_only after	editing_in_progress	history	contents
(1)	file_path is Some	unchanged	unchanged (false)	untouched	untouched
(2)	identity-protected (terminal)	unchanged (true)	unchanged	untouched	untouched
(3)	re-entrant on the same buffer	unchanged	unchanged (true, the outer edit's)	untouched	untouched
(4)	range out of bounds	unchanged	unchanged (false)	untouched	untouched
Rejected	mid-codepoint position, CRDT	restored to entry	false	untouched	untouched
AppliedThenFailed	a view rejected on_edit	true	false	cleared	mutated; carried Edit must fan out
Diverged	CRDT delete ok, insert failed	true	false	untouched	rope untouched, CRDT diverged
NoOp	empty write over an empty rope	true	false	cleared	unchanged
Applied	the ordinary case	true	false	cleared	replaced

Why end_edit cannot be skipped. It is line (6), unconditional and outside the match, in the same function as line (3) that set it — there is no caller who could return early past it, which is the whole reason the transaction is one Buffer method rather than a binding-level pair. Review round 2's P1-1 is right that a leaked editing_in_progress wedges the buffer for every later edit (begin_edit :726-731 and apply_edit :774-779 both refuse), and shipped set_generated_contents avoids that hazard today only by never setting the flag at all — which is also why it has no re-entrancy gate today, a gap this closes.

What bounds pre-validation is still for. It is no longer load-bearing for history (the Rejected variant covers that) but it is kept, for one reason: it moves the most common caller error out of the unlocked interval entirely, so an out-of-range op cannot even transiently unlock a buffer. It is not a substitute for the criteria — review round 3's P1-2 is exactly the trap of pinning the transaction through the one path that never enters it (§6, Stage 2 criteria 15-16).

Cost, stated rather than buried — and revision 4's outcome enum is the larger half of it. run_rope_edit_and_broadcast must distinguish Rejected from Diverged, and today it cannot: both surface as Err(CrdtRejected) because apply_to_crdt_then_normalize_bytes uses a bare ? on the second op (src/buffer.rs:1155-1163). Distinguishing them means that function tracking whether its delete succeeded before its insert failed. That is a real change to a shipped CRDT path, it is in Stage 2's scope, and it is the reason Diverged is a named variant rather than a note — a variant nothing can construct is not a design.

Revision 5 withdraws revision 4's four-variant fallback. Folding Diverged into Rejected would apply Rejected's cleanup: on a fresh writable buffer it restores entry_read_only = false, leaving a buffer whose CRDT and rope already disagree open to further writes. That directly contradicts the containment argument above.

Revision 6 keeps that withdrawal and widens the seam it rests on (round 5, P1-2). Revision 5 specified "a private delete→insert classifier that accepts the two operations as closures". Two operations is one too few: export_updates_since is a third fallible stage that runs after every successful CRDT mutation, so a two-closure seam cannot construct three of the seven cases in §3.4's table — the three where revision 5's classification was itself wrong. The seam is instead the crdt_mutated flag described in §3.4, and the injectable closures are the CRDT primitives and the export. Production supplies the loro calls; a #[cfg(feature = "crdt")] unit test supplies delete Ok + insert Err, and separately delete Ok + export Err. The test makes Diverged constructible without adding a public fault-injection API, and cargo test --lib --features crdt is its gate. If that extraction proves larger than expected, Stage 2 stops for review; it does not silently weaken the approved cleanup table.

validate_op_bounds is a new private helper duplicating the bounds arithmetic Rope::insert / delete / replace already perform (RopeError::OutOfBounds, src/rope.rs:371-383). It is O(1) and it is duplication; the alternative is a dry run, and there isn't one. Named in §7 as a bet.

---

4. Decisions

Reading order note: Q#GB15–18 are new in revision 3 and sit between Q#GB7 and Q#GB8 rather than at the end, because each descends directly from the decision above it — Q#GB15 and Q#GB16 are the two consequences of §2.11's finding that read_only is not this arc's flag, and Q#GB17 and Q#GB18 are what review P1-1 and P1-2 turned into decisions. The numbering is chronological; the placement is topical.

Q#GB1 — The streaming primitive. Buffer::apply_generated_edit(op), exposed as { generated = true } on the three Lua mutators. set_generated_contents becomes its whole-buffer wrapper, keeping name, Lua signature and behavioral tests; its Rust return becomes GeneratedOutcome so a failed-but-applied edit survives to the fan-out. Rationale and rejected alternatives: §3.

Q#GB2 — generated is additive; bypass_intercept stays. Seven call sites outside builtin/ depend on bypass_intercept, including tests/folding_stage2_acceptance.rs:1296-1315, which pins that a bypass edit still triggers the Q#FD19 interactive unfold. Redefining the existing key would silently change that pinned seam. generated = true implies bypass; passing both is legal and generated wins (it is strictly stronger); passing generated on a buffer with no intercept is legal (Class C would use it if it ever adopts).

Q#GB3 — A generated write gets its OWN run_buffer_edit arm. Reversed in revision 3 (review P1-1).

Revision 2 said "a generated write goes through run_buffer_edit's bypass arm". That is unimplementable — the bypass arm is run_bypass_edit, which calls begin_edit, which calls ensure_writable first (§3.4). run_buffer_edit (src/lua_bindings/mod.rs:1353-1374) grows a third arm and becomes the single owner of post-borrow fan-out. The three mutator bodies remove their separate notify_buffer_edit_to_windows calls, preventing the generated success arm from notifying twice:

if generated {
    // (round 5, P2-7) Refusals decide BEFORE the unfold, so a rejected
    // attempt cannot leave a visible side effect behind. See below.
    generated_preflight(lua, id, &op)?;
    unfold_before_interactive_lua_edit(lua, id, edit_start_of(&op));
    let outcome = run_generated_edit(lua, id, op); // no begin_edit; §3.4
    finish_generated_outcome(lua, id, outcome)     // fan-out, then Ok/Err
} else if bypass_intercept {
    unfold_before_interactive_lua_edit(lua, id, edit_start_of(&op));
    let edit = run_bypass_edit(lua, id, op)?;
    notify_buffer_edit_to_windows(lua, id, &edit);
    Ok(edit)
} else {
    let edit = run_managed_edit(lua, id, op)?;
    notify_buffer_edit_to_windows(lua, id, &edit);
    Ok(edit)
}

The refusal preflight, and why revision 5 needed one. New in revision 6 (round 5, P2-7). Revisions 3–5 unfolded before calling run_generated_edit, while all four refusals — path-backed, identity-protected, re-entrant, out-of-bounds — are decided inside apply_generated_edit, one registry borrow later. So an interactive generated attempt on a folded, file-backed buffer would open the fold and then report that nothing was touched. §3.4's contract says exit 1 is "before any state change"; a fold is state, and it is the state the user can see.

generated_preflight is a read-only with_registry borrow that evaluates §3.4's exits 1–4 against &Buffer and returns their errors verbatim. All four are decidable without mutation, so nothing is lost by asking early.

apply_generated_edit still re-checks all four, and that is not redundancy to remove. The borrow is released between the preflight and the apply, so the preflight is an optimization of the error path, not an authority. Buffer remains the only authority — which is the same reason §3.4 keeps the whole transaction in one method. The cost is that four cheap predicates run twice on the refusal path and once-plus-once on the success path; the alternative is an unfold that survives a refusal.

Considered and rejected: drop the unfold from the generated arm. That would make refusals safe by removing the behaviour entirely, but Q#GB3 keeps the unfold deliberately — an interactive M-x compile into a folded *compilation* should reveal what it wrote — and removing it would be a silent behaviour change to the pinned Q#FD19 seam. Narrowing the contract instead ("before any state change except folds") was also rejected: it makes the contract unfalsifiable exactly where a user can see it.

Note the asymmetry this leaves, deliberately. The bypass_intercept arm still unfolds before a write that can fail — but its only refusal is ensure_writable, which is pre-existing behaviour on a pre-existing path, and widening the preflight to it would change a shipped seam this lane does not own. Named rather than silently fixed.

run_generated_edit mirrors run_bypass_edit's borrow shape but returns the single transaction's whole outcome rather than calling begin_edit

• apply_edit_skip_intercepts. finish_generated_outcome is the match in §3.3 and is also used by the set_generated_contents binding. begin_edit stays byte-identical, so no ordinary edit's error precedence changes.

What revision 2 got right and revision 3 keeps: the unfold seam. The generated arm still calls unfold_before_interactive_lua_edit at the same point the bypass arm does, and for the same reason — the guard already requires InteractiveCommandOrigin::current() to be Some (src/lua_bindings/mod.rs:1424-1429), which is false for the process.after-tick pump and true for M-x compile, and the op is applied verbatim so the site is known up front (the round-5 F1 distinction the comment at :1359-1367 records). Keeping it is the no-change option.

But revision 2's stated reason for keeping it is now known to be half-false, and Q#GB16 is the consequence. Revision 2 wrote "folding a *compilation* buffer is possible, so changing this would be a silent behaviour change to a pinned seam". Folding a *compilation* buffer is possible only until this arc locks it: pmacs.fold.fold refuses every read_only buffer at src/lua_bindings/fold.rs:313. So the arc preserves the unfold-on-edit seam while silently killing the fold-creation seam that feeds it. The unfold arm stays because it costs nothing and because FoldRegistry::unfold_containing is registry-side and unaffected by the lock; the reason is corrected here so a reviewer does not inherit revision 2's version of it.

Q#GB4 — History cleared per op, with a measurement obligation. §3.3. Deferred optimization: suppress recording instead of clearing.

Q#GB5 — The lock-at-creation gap, and who closes it. A { generated = true } write locks the buffer after its first call, so between pmacs.buffer.create(name) and the owner's first generated write the rope is writable. dired.lua (claim_handle → paint), listview.lua (ensure_panel → render) and the search panel (ensure_search_panel → the header write in pmacs.project.search) all write synchronously in the same call, so the window is not observable. compile.lua's ensure_slot (:258-282) does not — it creates *compilation* and returns, leaving it empty and writable until start_run. Recommendation: ensure_slot ends with pmacs.buffer.set_generated_contents(slot.buf, ""), using the shipped primitive; no third surface is needed.

Amended in revision 2 (review P1-2), and the amendment is a hard ordering constraint, not a caveat. ensure_slot is buffer_named(name) or create (compile.lua:263), and pmacs.compile.run calls it before start_run validates opts.display (:1090 vs :752-757). §2.8 measures that a display = "bogus" call today raises and still leaves a foreign *compilation* permanently un-editable; with the empty write placed at the end of ensure_slot that same failing call would empty the buffer and lock the rope, unrecoverably. So the lock may only be installed once Q#GB13's ownership rule guarantees slot.buf is a buffer compile created. With ownership in place the buffer is provably fresh and the placement in ensure_slot is correct; without it, no placement is.

Revision 4: this recommendation is also the concrete caller that made P1-1 direction A load-bearing, and the dependency runs both ways. set_generated_contents(slot.buf, "") on a buffer that is already empty is a semantic no-op — it takes src/buffer.rs:1245-1253's early return. Under revision 3's revision-keyed cleanup that call would have locked the slot while leaving any pre-existing history intact and the buffer marked modified, which is the opposite of what "lock it at creation" is for. §3.4's NoOp variant clears unconditionally, so the recommendation is sound again. Stated here rather than only in §3.4 because a reader reaching Q#GB5 first should not have to derive it: the correctness of this placement depends on two decisions in other sections — Q#GB13's freshness guarantee and §3.4's NoOp cleanup — and criterion 16b is what pins the second.

Q#GB6 — Clamp each window coordinate against its OWN post-edit bound. §2.6 measures a shipped defect: a shrinking generated write leaves win.cursor past the end of the rope, and neither paint nor C-p recovers it. Recommendation: clamp in EditorCore::notify_buffer_edit — a clamp, not a call to rebuild_views_for, because a rebuild is O(buffer length) and would run per streaming op.

Revised in revision 2 (review P2-4). Revision 1 said "clamp when the buffer shrank", which conflates two different extents. The two coordinates are bounded by different things:

• win.cursor is a byte position (src/window.rs:366-367, "Byte position of this window's cursor"), bounded by Buffer::len().
• win.view_top is a line index (src/window.rs:373-374, "First buffer line shown at the top of this window's viewport"), bounded by TextView::line_count() (src/text_view.rs:67).

A replacement can grow in bytes while collapsing many lines into one — "a\nb\nc\nd\ne\nf\n" (12 bytes, 7 lines) replaced by a single 80-byte line — leaving view_top invalid on a write that a byte-length comparison calls a growth. So the trigger cannot be "the buffer shrank": the clamp runs unconditionally, each coordinate against its own bound, exactly as rebuild_views_for already does (src/editor_core.rs:1875-1879, which clamps cursor against len and view_top against line_count().saturating_sub(1)).

Argued from the types and from rebuild_views_for's existing shape, not measured — unlike §2.6's cursor case, the view_top case needs a scrolled window to stage and was not staged. §6 Stage 1 criterion 8b is what turns the argument into a pin.

A THIRD window coordinate, and it crashes rather than dangles. New in revision 6 (round 5, P1-5). Every window also owns an optional Selection, and Selection::anchor is a byte position (src/window.rs:120-130, "Where the selection began"). Revisions 1–5 listed two coordinates and there are three.

The anchor is worse than the other two because nothing downstream tolerates it being out of range. EditorCore::region_bytes (src/editor_core.rs:4184-4191) does:

let (lo, hi) = self.active_region()?;              // Window::region(), :472-479
let mut out = vec![0u8; (hi - lo) as usize];
buf.snapshot_rope().slice(lo, hi, &mut out);

and Rope::slice asserts its bounds — debug_assert!(end <= self.len()) at src/rope.rs:145. Window::region (src/window.rs:472-479) returns (anchor, cursor) in canonical order with no clamping of either. So a stale anchor is not a cosmetic dangle: #191 reproduced the crash — select bytes 0..30 in a generated buffer, let the owner rewrite it to two bytes, press copy, panic at src/rope.rs:145. A shrinking refresh under a live selection is not an exotic sequence; it is g on a listview panel with a region marked.

The rule: clamp-or-clear, per window, in BOTH functions. For every window whose buffer_id matches — the windows.values_mut() loop that both functions already run:

1. win.cursor → clamp to len.
2. win.view_top → clamp to line_count().saturating_sub(1).
3. win.selection → clamp anchor to len; then, if the clamp collapsed the selection — anchor == cursor afterwards and at least one endpoint actually moved — set selection = None.

Rule 3 is not invented here; it is the tree's own answer to the same question. src/terminal/view.rs:715-721 normalizes the terminal's selection against a shrinking scrollback and does exactly this:

state.selection = state.selection.and_then(|selection| {
    let anchor = clamp_or_clear(&rows, selection.anchor)?;
    let head = clamp_or_clear(&rows, selection.head)?;
    let collapsed_by_clamp =
        anchor == head && (anchor != selection.anchor || head != selection.head);
    (!collapsed_by_clamp).then_some(TerminalSelection { anchor, head })
});

The terminal subsystem solved this for TerminalSelection and the window Selection never got the same treatment. Adopting the same rule rather than a fresh one is the point: a bare clamp would leave a zero-width "active but empty" selection that Selection's own doc says is legal (src/window.rs:123-126) but that the user never asked for, and collapsed_by_clamp is precisely the distinction between "the selection survived, shortened" and "the selection's content is gone".

Both functions, because they have different callers and neither subsumes the other:

• notify_buffer_edit (src/editor_core.rs:1836-1850) clamps nothing today — it only forwards on_edit to text_view and overlays. This is the function a generated write reaches, so this is where the crash is fixed.
• rebuild_views_for (src/editor_core.rs:1865-1882) already clamps cursor and view_top and does not touch selection. It serves the *help* and *buffer-list* rewrites, which are Class C and shrink wholesale. The same crash is reachable there today, independently of this arc.

Blast radius of a stale anchor beyond the crash, so the fix is not undersold: Window::region also feeds the presence broadcast (src/presence.rs:122-123, SelectionSnapshot), so an out-of-range anchor is published to peers as well as sliced locally.

Recommended for Stage 1, because Stage 1's adopters refresh shrinking panels constantly and because it fixes terminal copy mode retroactively. The selection half is not optional within that: it is the only one of the three that panics, and it is reproducible today. Alternative if the user prefers a narrower Stage 1: its own lane, in which case Stage 1 must say so out loud rather than inherit it silently.

Cross-lane note. Per the boundary in §9b, this rule is #188's to specify and #191's to implement, and both must describe the same rule. The text above is the specification; #191 adopts it rather than restating it.

Q#GB7 — No unlock ships in this arc. Revision 5 chooses the fallback revision 4 named (review round 4, P1-4).

Revision 2 recommended pmacs.buffer.unlock_generated(buf) — a one-way clear of read_only, shipped in Stage 1 — on the strength of sweep B's finding that the two halves of the protection are asymmetric: the intercept half is removable (remove_intercept, src/lua_bindings/mod.rs:3433, used by the REPL at repl/init.lua:325-327) while the rope half is one-way from Lua. That observation stands. The capability revision 2 derived from it does not, on two independent counts, both of which the review is right about.

First: an unbounded clear of read_only is not a generated-buffer capability at all. §2.11 measures that the flag serves three unrelated policies. unlock_generated as revision 2 wrote it — "clears read_only and nothing else" — would let any Lua caller disable a live terminal identity buffer's protection (src/terminal/session.rs:305), which this arc never locked, whose owner set it to refuse host edits, undo, redo and remote CRDT imports alike, and which is Lua-reachable (pmacs.terminal.open returns its id; pmacs.terminal.is_terminal(buf) exists at src/lua_bindings/mod.rs:8858-8867). That is a strictly larger capability than the one the arc adds, granted by accident.

Second: it does not achieve its own stated purpose. Revision 2 sold it as the escape from an accidentally bricked *scratch*. It is not: by the time anyone reaches for it, set_generated_contents has already replaced the contents and cleared the history (§3.4). Unlocking returns writability to a buffer whose data is gone. The recovery scenario that justified moving this from a deferral to Stage 1 work was never a recovery.

Revision 4's proposed bound was not provenance. It refused only when identity_protected was true. That proves “this is not a declared terminal identity buffer”; it does not prove “the generated-write primitive installed this lock.” As revision 4 itself admitted, unlock_generated could therefore release any non-identity-protected Rust lock, including one a future owner installed for an unrelated reason. The live Q#GB7 text simultaneously promised to refuse exactly those locks. Both statements cannot be the approved contract.

The decision: remove the capability, not the promise. Neither stage adds pmacs.buffer.unlock_generated, and Q#GB15's field is used only to refuse generated writes to intrinsic identity buffers. There is no Lua clear of read_only.

This leaves the standing asymmetry visible rather than pretending to close it: remove_intercept is exposed while set_read_only is not (src/lua_bindings/mod.rs:3072-3078 and docs/agent-handoff.md §4). That asymmetry is deliberate today because the rope lock protects more than generated buffers. A safe inverse needs a lock-kind/provenance representation, or an owner-specific transition whose preconditions prove the caller owns the state. Neither is required to stop undo destroying generated output, and neither should be smuggled into this arc for an escape that cannot restore the overwritten data.

The future concrete consumer remains wdired in dired Stage 3. That lane must frame the transition it needs—generated listing to editable rename surface—against its own owner handle and lifecycle. It may choose the eventual read_only policy enum; it does not inherit a general-purpose unlock from here.

What Stage 1 loses, and why that is correct. Nothing. Stage 1 adopts set_generated_contents, which is already public on main, on two more buffers. It therefore adds no brick capability the tree does not already ship, and revision 2's argument that Stage 1 needed an escape hatch applied equally to main — which is a sign the argument was about the shipped primitive, not about Stage 1. The *scratch* exposure sweep B found is real, and it is real today; it is recorded in §8 as a pre-existing hazard this arc neither creates nor closes.

Q#GB15 — read_only gains an intrinsic-identity guard, not a provenance companion. New in revision 3, narrowed in revision 5 (review P1-3, sweep C; review round 4 P1-4).

Revision 4 replaces revision 3's generated_lock with an intrinsic identity_protected, because review round 3's P1-3 showed provenance cannot be inferred from the flag's history.

Revision 3 proposed generated_lock: bool, meaning "this buffer's read_only was set by a generated write", maintained by a third rule: every set_read_only call clears it. That rule breaks a seam this very document prescribes. tests/terminal_copy_mode_acceptance.rs:578-584 does exactly this cycle, with a comment explaining why:

// `read_only` refuses the upgrade's own bookkeeping path the same
// way it refuses everything else, so lift it around the upgrade.
buffer.set_read_only(false);
buffer.upgrade_to_crdt(2).expect("upgrade");
buffer.set_read_only(true);

Under revision 3's rule the final true yields read_only && !generated_lock, so the next owner refresh of that snapshot is refused as someone else's lock. Q#GB12 and Stage 2 criterion 4 prescribe the same lift-and-restore idiom, so revision 3 invalidated its own test strategy. The defect is not the rule's details; it is that a derived fact has to be maintained correctly by every mutation of the thing it is derived from, and set_read_only is pub with callers this document does not control.

The fix is to stop deriving it. Buffer gains one private field that is a property of what the buffer is, not of who last locked it:

/// Whether this buffer's read-only state is an intrinsic identity
/// protection rather than an ordinary lock. Marked once during owner setup,
/// by an owner that means "the host may not edit this at all"; never
/// derived from `read_only` and never changed by `set_read_only`.
identity_protected: bool,

Two rules maintain it, and set_read_only is not one of them:

1. Buffer::mark_identity_protected() — a new crate-private, monotonic method that can only set the field to true, called once by TerminalSession::open beside its existing set_read_only(true) (src/terminal/session.rs:305). There is no “false” operation and no Lua surface, so “set once” is enforced rather than a caller convention.
2. apply_generated_edit refuses iff identity_protected (§3.4 exit 2). It never writes the field.

The lift-and-restore seam is now unaffected, because identity_protected is false for a snapshot buffer and stays false through any number of set_read_only cycles. Q#GB12's prescription and Stage 2 criterion 4 need no change, which is the test that the rule is right.

Why declaration beats inference, stated as a rule rather than as a patch. Inference required every mutation of read_only to maintain a derived fact; P1-3 is the proof that it does not. Declaration is sound for the one question this field now answers: may an owner-authorized generated write ever edit this buffer? Terminal identity says no, independently of temporary read_only lifts. It is deliberately not used to answer who installed an ordinary lock; Q#GB7 no longer asks it to.

Rule 1's refusal is the half revision 2 did not have, and it closes a hole in the SHIPPED primitive (sweep C item 1). Buffer::set_generated_contents today does self.read_only = false unconditionally (src/buffer.rs:546), so pmacs.buffer.set_generated_contents(term_buf, "junk") on a live terminal identity buffer overwrites its contents and re-locks it as though the primitive owned it. Nothing in the tree refuses that, and no test covers it. That write-direction hole is the field's sole purpose.

The alternatives, and why not.

• A registry-side set of generated-locked ids, held as Lua app-data. This would exist only to resurrect Q#GB7: a second source of truth that can drift from the flag, plus a pruning obligation on buffer removal — the shape the terminal-config lane records as "prune reacts to buffer removal". Rejected along with the unlock.
• Replacing read_only: bool with an enum. Cleaner in principle, and it would let §2.11's third policy (document_bytes) ask the question it actually means. It also churns ensure_writable, all six gated paths, is_read_only's seven callers and the public set_read_only signature — a refactor this arc would be smuggling. Named as the right eventual shape in §8, not adopted.

Cost, stated: one bool per buffer; one new crate-private monotonic method with exactly one production caller; no invariant to maintain against set_read_only, because the field is never derived from it; and one new refusal that changes shipped set_generated_contents behaviour, so it lands in Stage 2 with the rest of Q#GB10's changes to that function, not in Stage 1.

Q#GB16 — The lock silently disables fold creation on every buffer it touches. New in revision 3 (sweep C item 2).

document_bytes (src/lua_bindings/fold.rs:310-318) is Q#FD11's "normal document buffer" guard and it is spelled if buffer.is_read_only() { return Ok(None); }. Its two consumers are pmacs.fold.fold (:67-70, which then sets the status fold rejected: not a document buffer and returns false) and pmacs.fold.unfold's normalize-an-arbitrary-range fallback (:108). tests/folding_acceptance.rs:570-590 (read_only_buffer_is_rejected) pins the behaviour, and its comment records the intent the guard was written with: "terminals are read-only, so a read-only buffer is not foldable."

So the moment Stage 1 locks dired listings and listview panels, and Stage 2 locks *compilation*, *shell-command* and *search-results*, pmacs.fold.fold starts answering false on all five families — with a status message that is now false ("not a document buffer"), on a seam nothing in this arc's acceptance would notice, and against a guard whose author meant "terminal", not "generated".

Recommendation: name it, do not silently accept it, and do not fix it here. Three options, with the recommendation being (a):

• (a) Accept, and pin the acceptance. A generated buffer arguably should not be foldable — its contents are replaced wholesale and any stored fold range is invalidated on every refresh anyway. Then the change is intended, and Stage 1 owes an explicit criterion asserting it, plus the status string corrected to say read-only, not "not a document buffer". Cheap, honest, and it converts a silent behaviour change into a stated one.
• (b) Preserve foldability by changing the guard to the intrinsic identity_protected predicate rather than the broad read_only predicate. Available once Q#GB15 lands, but it edits a pinned Q#FD11 seam for a use case nobody has asked for.
• (c) Do nothing and say nothing. Rejected: this is exactly the defect class the review's findings 1 and 3 are instances of.

Under (a) the only code change is the status string; the substance is the criterion. Stage 1, because Stage 1 is where the first three families get locked.

Q#GB17 — The transaction shape. §3.4. One &mut Buffer method, its own run_buffer_edit arm, begin_edit untouched, eight named exits, and cleanup driven by an explicit GeneratedOutcome — not by inferring one from revision, which revision 3 did and which was wrong in three directions (§3.4). New in revision 3 (review P1-1).

Q#GB18 — Route the two broken identity consumers by owned BufferId. New in revision 3 (review P1-2).

§2.10's census finds exactly two sites that a disambiguated name breaks, with six downstream consumers between them. Q#GB13 removes the adoption; this removes the recognition that adoption was hiding.

listview.lua (Stage 1, with Q#GB13's ownership fix). panels becomes a list of records rather than a name-keyed map, with two lookups instead of one:

• panel_for_requested_name(name) — matches p.requested_name, the spec.name the caller asked for. Stable across disambiguation, so repeated listview.open{ name = "*references*" } finds the same panel.
• panel_for_buffer(buf) — scans p.buffer == buf. This is what listview.visit, listview.refresh, listview.quit and listview.open's capture guard use.

The list-and-scan shape is deliberate and is dired's, for the reason dired.lua:123-127 already records: two BufferIdLua values for the same buffer are distinct userdata, so panels[buf] would miss even for the same buffer; handle_for_buffer (dired.lua:142-148) compares with == instead, and compile.lua's slot_for_buffer (:200-206) does the same. Three existing implementations of one shape; listview adopts it rather than inventing a fourth.

And listview needs dired's other half too, which is easy to miss. grep -n on_removed builtin/runtime/listview.lua returns nothing — listview registers no buffer-removal callback, unlike compile (:277-279) and the search panel (default.lua:876-883). Today that is harmless because ensure_panel re-checks p.buffer:is_valid() on every open (:93-94) and a name-keyed map holds at most one entry per name. A list does not self-limit: kill and reopen *references* ten times and a naive list holds ten records, nine of them dead, and panel_for_buffer's scan walks all of them. So the list must compact on scan, exactly as dired.lua:132-140's live_handles() does. Naming it here because "swap a map for a list" reads like a one-line change and is not.

compile.lua and the search panel (Stage 2, with Q#GB13's ownership fix). Rewritten in revision 4 (review round 3, P2-4).

is_generated_buffer stops comparing names. Its two owners are in different files (compile.lua's slots, default.lua's search_panel), so the predicate needs a seam — and revision 3 chose the wrong one.

Revision 3 had ensure_search_panel call pmacs.compile._register_generated_buffer(p.buf). That converts an existing guarded, optional dependency into a hard one. default.lua:991-994 currently reads:

if cur
  and not (pmacs.compile
    and pmacs.compile.is_generated_buffer
    and pmacs.compile.is_generated_buffer(cur))
then

— a triple check, and it exists for a reason. default.lua is loaded by LuaHost::attach_editor (src/lua.rs:250-251), while compile.lua is loaded much later in EditorState::new's runtime sequence (src/editor.rs:704). LuaHost is pub and nine existing test files build one directly, so a harness that has pmacs.project.search and no pmacs.compile is not hypothetical — it is reachable with the pattern those nine already use. A hard call to pmacs.compile._register_… inside ensure_search_panel raises there.

The decision: symmetric guarded optionality, not a shared registry. Each module answers for its own buffers, and each capture site ORs the two through the guard shape that already exists:

• compile.lua — is_generated_buffer(buf) becomes slot_for_buffer(buf) ~= nil, which is already id-based (:200-206). It answers for *compilation* and *shell-command* only, which is the truthful scope and matches §1.4's warning not to read ensure_slot as covering the grep panel.
• default.lua — a local search_panel_owns(buf), i.e. search_panel ~= nil and search_panel.buf == buf, exposed as pmacs.project._is_search_panel for the other direction.
• Each capture site checks its own predicate directly and the other module's through the existing optional guard:
  • default.lua:991-994 keeps its triple check verbatim and gains and not search_panel_owns(cur);
  • compile.lua:762 gains the mirror-image guard, and not (pmacs.project and pmacs.project._is_search_panel and pmacs.project._is_search_panel(cur)).

Three things this buys over revision 3's registry, all of them consequences of nobody owning a list of somebody else's buffers:

1. No load-order constraint in either direction, and no new one to document beside the three src/editor.rs already carries.
2. No teardown obligation. Revision 3's registry needed on_removed-driven unregistration, and a registry that only grows is the defect the terminal-config lane records as "prune reacts to buffer removal". Each owner's own table already tracks its own liveness (slot_for_buffer scans slots, search_panel_owns reads one field), so there is nothing to prune.
3. Each predicate is answerable by the module that knows the answer, so neither can go stale relative to the other.

Cost: the guard shape is written twice rather than once. That is the existing pattern, and duplicating a four-line guard is cheaper than a cross-module registry with a lifetime.

This needs its own pin, because no ordinary test reaches it — every EditorState::new loads both modules. §6 Stage 2 criterion 21 builds a LuaHost, calls attach_editor, and drives search with no pmacs.compile present.

Alternative considered and rejected: leave is_generated_buffer name-based and simply never disambiguate compile's buffers. Rejected because Q#GB13's whole argument is that a foreign *compilation* must not be adopted, and refusing to adopt without disambiguating means raising — which turns a name collision into a failed M-x compile.

Q#GB8 — The REPL is out of this arc. §2.5. Same root cause, different remedy, its own lane. Its measured exposure is recorded above so the next scout starts from evidence.

Q#GB9 — Class C is out of this arc. *buffer-list*, *help* and *workers* are generated but were never claimed to be protected, so nothing about them is defeated. Making them immutable is a product decision about COHERENCE.md §14's list and output-channel primitives, not a bug fix, and it should not ride a bug-fix arc. *workers* additionally writes from Rust with its own fan-out pair and already mark_cleans, so it is not a like-for-like conversion.

Q#GB10 — Refuse a generated write on a path-backed buffer; then, and only then, mark clean. Rewritten in revision 2 (review P1-3).

set_generated_contents leaves is_modified = true (measured), so every adopter shows * in the mode line (src/editor.rs:4129) and in *buffer-list* (default.lua:395). workers_buffer::render calls Buffer::mark_clean() (src/workers_buffer.rs:95) and instance_buffer.rs:401 asserts the same for its own rendered buffer, so marking clean is the established convention for a generated buffer.

Revision 1's justification was wrong. It said the flag "drives only the mode-line indicator and the buffer-list column". §2.9 measures two more consumers: src/autosave.rs:363, the skip that decides whether a crash-recovery slot is written, and src/desktop.rs:302. Since { generated = true } is public Lua on any buffer id, a caller could replace a file-backed buffer's contents, mark it clean, and suppress autosave recovery for it.

The rule, stated explicitly rather than left implicit: Buffer::apply_generated_edit (and therefore set_generated_contents) returns an error for a buffer whose file_path() is Some. Then mark_clean is unconditionally safe, because both consumers gate on file_path() before they read the flag (autosave.rs:359-364, desktop.rs:298-303).

Why refuse rather than the alternative "retain modified state for path-backed buffers": the flag rule fixes only the flag. A generated write on a file buffer would still replace its contents and lock its rope, and §1.3 measured that Lua cannot unlock. Refusing bounds all three harms with one rule, and it is the narrower capability.

Verified non-breaking. None of the six generated families is path-backed: they are all pmacs.buffer.created, and no builtin Lua sets a buffer path — grep -rn "set_path\|set_buffer_path" builtin finds no call sites (only a comment in dired.lua:42 and lsp.lua's own active_buffer_path local). Path binding happens Rust-side in from_file / find_file only.

This changes shipped set_generated_contents behaviour — both the new refusal and mark_clean — and therefore the terminal snapshot, so it belongs in Stage 2 alongside the reimplementation, not smuggled into Stage 1.

Q#GB11 — Staging. §5.

Q#GB12 — The revision guard becomes near-dead, and three tests break. After Stage 2, an external edit to *compilation* or *search-results* is refused at the rope, so the Q#CM2 desync machinery (check_rev compile.lua:332, resync :309, search_panel_check_rev default.lua:832) can essentially no longer fire. Recommendation: keep it — it is cheap, and a future Rust-side writer could still mutate the buffer — but say so, and do not delete its tests. Three compile acceptance tests inject intruder edits through bypass_intercept (tests/compile_mode_acceptance.rs:1040, :1106, :1305) and will be refused after conversion; they must lift read_only Rust-side first, exactly as tests/terminal_copy_mode_acceptance.rs:582-584 already does. That is a concrete, verified integration cost of Stage 2, not a surprise to discover during implementation.

Revision 4: the lift-and-restore idiom this prescribes was briefly invalidated by revision 3's own Q#GB15, and is now safe again. Revision 3's generated_lock was cleared by every set_read_only call, so the restoring set_read_only(true) would have left the buffer looking like someone else's lock and the next owner refresh would have been refused (review round 3, P1-3). Q#GB15's identity_protected is never written by set_read_only, so the cycle is transparent and this prescription, Stage 2 criterion 4 and acc16e all need no change. The seam is crdt-gated, so §10 now names --test terminal_copy_mode_acceptance --features crdt explicitly: a default-feature sweep never compiles acc16e and proves nothing about it.

Q#GB13 — Ownership by handle is a prerequisite, not a follow-up. New in revision 2 (review P1-2). listview.ensure_panel (listview.lua:95), compile.ensure_slot (compile.lua:263) and ensure_search_panel (default.lua:861-868) adopt any buffer that shares their name. §2.8 measures the consequence today (a clobbered, permanently un-editable user buffer — and, for compile, from a call that raised), and measures that M-x buffer.undo is currently the only recovery. Locking the rope removes that recovery, so the rule must land in the same stage as the lock.

Recommendation: adopt the rule the tree already states at terminal.lua:300-305 and implements at dired.lua:476-504 — ownership means "this buffer is in my handle table", a name collision disambiguates <2>…<99>, and exhausting the limit raises rather than adopting. Three writers, one shape, each in the stage that locks it: listview in Stage 1, compile and search in Stage 2. Dired and terminal already comply.

Alternative considered and rejected: a standalone Stage 0 that fixes all three at once. Rejected because each writer's ownership fix is only load-bearing for the stage that locks that writer, and a lone ownership PR reads as unmotivated churn without the lock that makes it urgent. If the user prefers the standalone shape, the acceptance criteria in §6 move with it unchanged.

Revision 3: Q#GB13 is only half the work, and revision 2 shipped the half that is visible. Disambiguating a name is not free — it changes what a buffer is called, and five sites in builtin/ recover identity from what a buffer is called. Two of them break (§2.10 Class 1) and are Q#GB18's. Stated as an ordering constraint so it cannot be split across PRs: within each stage, Q#GB18's routing change must land in the same PR as Q#GB13's disambiguation for the same writer. Disambiguate first and you ship a listview panel whose RET, g and q are dead and whose q-target capture is inverted; route first and you have written a lookup nothing yet exercises.

Q#GB14 — The lock is not observable from Lua, and the pins depend on it. New in revision 2, out of P1-1's fix. describe.buffer returns name, length, modified, view_count and nothing else (buffer_info_table, src/lua_bindings/mod.rs:6352-6364), so no Lua assertion can read read_only directly. Two discriminators are available and both are used in §6: a bypass_intercept write, which lands on main and raises buffer `X` (id BufferId(n)) is read-only once the rope is locked (measured, §2.4), and Rust-side Buffer::is_read_only() (src/buffer.rs:494, already pub). Recommendation: use both, and do not add a Lua surface for it — the acceptance suites are Rust and need no new public API. Optional and separable: adding read_only to buffer_info_table would be a read-only introspection field with no new capability, useful if Lua-level pins are ever wanted; it is not required by this arc.

---

5. Staging

The proposed cut is endorsed, with two amendments. The argument for it is not the obvious one.

Stage 1 — generated-buffer-immutability-stage1

dired.lua and listview.lua adopt pmacs.buffer.set_generated_contents.

• Prerequisite, in this PR, before the lock (Q#GB13): listview.ensure_panel (listview.lua:95) stops adopting a same-named foreign buffer. Ownership is the handle table (panels); a name collision disambiguates <2>…<99> and raises at the limit, matching dired.lua:486-504. dired.lua needs no ownership work — it already complies, which is why it is the cheaper of the two adopters.
• In the SAME PR as that disambiguation (Q#GB18): panels becomes a compacting list; panel_for_current_buffer is replaced by panel_for_buffer(buf) scanning p.buffer == buf, and ensure_panel looks up by p.requested_name. Four consumers move with it, including listview.open's capture guard (:118-123). Ordering constraint, not a preference — see Q#GB13.
• listview.lua:50-62 — render's delete-all + insert-all becomes one set_generated_contents(buf, body).
• dired.lua:369-372 — paint's whole-buffer replace becomes one set_generated_contents(handle.buf, text).
• Both keep their erroring intercept (named error, per the layering at terminal.lua:351-366) and both keep set_round_trip_input.
• Plus Q#GB6's per-coordinate clamp, if approved.
• Plus Q#GB16's fold decision — under recommendation (a), the fold.rs:68 status string, and the criterion that makes the change stated rather than silent.

Revision 3 removed unlock_generated from Stage 1; revision 5 removes it from the arc (Q#GB7). Revision 2 put it here as "the escape from a bricked buffer"; the escape does not recover anything (the history is already cleared), and the brick it escapes is one main already ships, since set_generated_contents is already public. Stage 1 adds no lock capability that does not already exist, so it needs no door. Revision 4's attempt to restore the capability in Stage 2 used identity_protected as though “not terminal identity” proved “generated lock”; it does not. No stage adds the binding.

Revision 2 grew Stage 1 by two prerequisites and one reversal. The ownership rule is load-bearing for the lock rather than adjacent to it, and it stays. The unlock does not, per the paragraph above. Stage 1 is still not a pure-Lua change — Q#GB6's clamp is Rust, and under Q#GB16(a) so is a status string — so revision 1's "pure Lua" claim stays withdrawn.

Why this cut, and why Stage 1 is not merely "the cheap half": it is the worse-exposure half. compile.lua:219 and builtin/commands/default.lua:855 rebind all seven undo chords to compile.undo-noop; dired.lua and listview.lua rebind nothing — grep -n 'C-/\|C-_\|C-x u\|undo' builtin/runtime/dired.lua builtin/runtime/listview.lua returns zero binding lines. Measured, a bare C-/ empties a listview panel and a dired listing. Stage 1 closes the only two families reachable without M-x.

Why the cut is safe under every candidate primitive: a whole-buffer replace is expressible in all of A–C, and under the recommendation set_generated_contents keeps its Lua name and signature as apply_generated_edit's wrapper. Stage 1 is therefore not rework under any Q#GB1 outcome — which is the decisive argument for cutting here rather than shipping one large PR.

The honest objection, and the answer. A reviewer could call Stage 1 churn: two call sites converted to a primitive Stage 2 then rewrites. Stage 2 rewrites the primitive's implementation, not its callers; the diff at dired.lua:371 and listview.lua:60-61 is written once.

Stage 2 — generated-buffer-immutability-stage2

• Prerequisite, in this PR, before the lock (Q#GB13): compile.ensure_slot (compile.lua:263) and ensure_search_panel (default.lua:861-868) stop adopting same-named foreign buffers, same shape as Stage 1's listview fix.
• In the SAME PR as that disambiguation (Q#GB18): pmacs.compile.is_generated_buffer (compile.lua:212-217) stops comparing names and answers only slot_for_buffer(buf) ~= nil; default.lua adds local search_panel_owns(buf) and exposes pmacs.project._is_search_panel. Each capture site ORs its own predicate with the other module's predicate through the guarded optional call shape Q#GB18 specifies. No owner registers ids with the other and no new removal callback work exists.
• Buffer::apply_generated_edit (§3.4) + the { generated = true } option + its own run_buffer_edit arm + set_generated_contents reimplemented over it (Q#GB17, Q#GB3).
• Q#GB10's path-backed refusal and mark_clean — one rule, both halves, since the refusal is what makes the flag change safe.
• Q#GB15's identity_protected field and its write-direction refusal. mark_identity_protected() is crate-private and monotonic; TerminalSession::open is its only production caller. Q#GB7 adds no unlock surface.
• Conversion of all 13 remaining write sites (compile.lua 9, builtin/commands/default.lua 4).
• Q#GB5's ensure_slot lock, which is only placeable once ownership lands.
• The three compile_mode_acceptance intruder tests updated per Q#GB12.

All the new Rust and all the review risk in one PR, which is the point of the cut.

Amendments to the briefed cut:

1. Q#GB13 (ownership) is a prerequisite of the stage that locks each writer, not a follow-up and not a separate PR. §2.8 is the argument: this arc removes the only recovery a clobbered buffer currently has.
2. Q#GB18 (identity routing) rides in the same PR as Q#GB13 for the same writer, per the ordering constraint under Q#GB13. New in revision 3.
3. Q#GB7 (unlock) is removed from both stages. Revision 1 deferred it; revision 2 built it in Stage 1; revision 3 moved it to Stage 2; revision 4 replaced its provenance with an identity exclusion and thereby made it general again. Revision 5 chooses no unlock. The reversals are recorded rather than smoothed over because the reason moved three times and the next reader needs to know which decision is live.
4. Q#GB10 (path refusal + mark_clean) lands in Stage 2, because it edits set_generated_contents itself and therefore changes the already-shipped terminal snapshot. Stage 1 is not pure Lua (Q#GB6's clamp, and Q#GB16's status string), which revision 1 claimed and revision 2 withdrew; the withdrawal stands for a different reason than revision 2 gave.

Where the REPL lands: neither stage. Q#GB8.

---

6. Acceptance, with the pre-image each criterion must fail against

M-x buffer.undo is the user-reachable trigger and needs no keymap. A criterion that only exercises the intercept, or only the chords, proves nothing — that is precisely what compile.lua's idiom already achieves and what this bug already defeats.

Revision 2 re-audited every criterion, not only the three the review named (sweep A). Each now carries an explicit pre-image class, because an unlabelled always-green criterion is indistinguishable from a vacuous one:

class	meaning
main	fails on ad41cf1. A regression pin in the ordinary sense.
fix-shape	passes on main by design; fails against a specific wrong implementation, named in the criterion. Legitimate per docs/agent-handoff.md §5 ("bite against every pre-image the fix could plausibly have taken"), where acc 6 deliberately passes on main.
mutation	passes on main; fails against a named one-line mutation of the fix.
structural	no behavioural pre-image. Rides alongside the others, never instead — a structural comparison of two authorities does not catch a misrouted consumer.

Q#GB14: the lock is not observable from Lua. describe.buffer carries no read_only field, so every "is it locked" assertion below uses a bypass_intercept write (lands on main, raises buffer `X` (id BufferId(n)) is read-only once locked) or Rust-side Buffer::is_read_only(). An ordinary edit is not a discriminator: the intercept refuses it either way.

Stage 1

1. [main] C-/ cannot empty a listview panel. Driven by dispatch_key, not a Lua call. Bite: measured — "H\nrow-one\nrow-two" → "".

2. [main] M-x buffer.undo cannot empty a listview panel, driven through the real minibuffer (M-x, type buffer.undo, RET), not pmacs.command.invoke. Bite: same empty result; and a chord-only fix passes 1 and fails this.

3. [main] C-/ and M-x buffer.undo cannot empty a dired listing. Bite: measured — one undo takes the listing to "".

4. [fix-shape] The owner's own refresh still works after the lock — g on a listview panel and on a dired buffer renders new content. Bite: a naive set_read_only(true) at creation passes 1–3 and fails here; that is the failure mode src/buffer.rs:521-524 exists to prevent. Assert the new content appears, not that the call did not raise.

5. An ordinary edit is refused, and the intercept is still installed and still fires. REWRITTEN in revision 6 (round 5, P1-1) — the old criterion was impossible, and this document is where that gets settled, not #191.

   Why the old one could not pass. It required the intercept's message on an ordinary edit after adoption. Both entry points check the lock before any intercept runs: begin_edit (src/buffer.rs:724) and apply_edit (src/buffer.rs:772) each call ensure_writable() as their first statement, and the chain runs later, in apply_edit_inner. Once the rope is locked an ordinary edit necessarily returns BufferError::ReadOnly; the intercept cannot run, so no test can observe its text. #191 reached this independently and measured the actual message.

   The consequence revisions 1–5 all missed: after adoption the intercept is UNREACHABLE on the ordinary path — including in the shipped precedent. terminal.lua's claim_snapshot keeps its erroring intercept beside the rope lock (:339-396), and that intercept has been dead since #178 landed. This document has been telling two more adopters to preserve it and to assert its message.

   So why keep the intercept at all? Because it is the guard that remains whenever the lock is lifted, and this document prescribes lifting it: Q#GB12's intruder-test conversion, criteria 4, 15, 16b and 17's Rust-side lifts, and — if dired Stage 3 ever ships a wdired mode swap — the editable window. During any such window the intercept is the only thing refusing an ordinary edit. That is a real role, and it is testable.

   The criterion, in two halves:

   • (a) [main] An ordinary edit is refused and the text is byte-identical. No claim about which guard refused. Driven through dispatch_key.
   • (b) [fix-shape] With the lock lifted Rust-side, an ordinary edit is refused BY THE INTERCEPT, asserted on the message text: intercept rejected the edit: … is read-only, not buffer `X` (id BufferId(n)) is read-only. Restore the lock afterwards.

   Bite: (b) is where the old criterion's bite survives — an adopter that deletes the intercept and relies on the rope alone passes 1–4 and (a), and fails (b). (a) alone is not a discriminator, which is exactly why it is split out rather than left to carry the claim. The layering at terminal.lua:351-366 is preserved by (b); what revisions 1–5 got wrong was believing the ordinary path could see it.

   Recorded consequence, not this arc's to fix: terminal.lua's intercept is likewise reachable only under a lift. Whether a permanently-unreachable-on-the-ordinary-path guard should stay is a question for whoever owns the layering; §8 carries it.

6. [fix-shape] set_round_trip_input is still set on both — asserted so that only the round-trip mark can make it pass. Rewritten in revision 3 (review P2-4), and the cited precedent was wrong.

   dispatch_idle_for (src/editor.rs:1126-1155) returns false for six independent reasons, only one of which is the round-trip mark: a pending chord or terminal escape on that frontend (:1130), an active minibuffer, an active search, an active query-replace, an open menu (:1135-1138), a focused side window and core.buffer_round_trips(window.buffer_id) — both on :1153. A criterion that only asserts !dispatch_idle_for(..) is satisfied by any of the six.

   Revision 2 named the wrong model. It cited tests/terminal_copy_mode_acceptance.rs "criterion 16", but that file contains zero dispatch_idle_for references — acc16 (:321-339) goes through state.dispatch_idle() and asserts no is_side premise. The test that gets the side-window half right is tests/dired_acceptance.rs:969-1013, which asserts !window.is_side() as an explicit fixture premise (:975-989, commented "A document window, deliberately: the panel arm of the same gate would otherwise be what makes this pass") before asserting !s.dispatch_idle_for(FrontendId::LOCAL).

   Both halves are required, because neither test has both. The criterion asserts, for each of the dired and listview adopters:

   • (a) the document-window premise, !window.is_side() on the focused window, asserted as a premise so a fixture that later displays in a panel fails loudly rather than passing vacuously — dired_acceptance.rs:975-989's shape verbatim;
   • (b) !dispatch_idle_for(FrontendId::LOCAL) while the panel is focused;
   • (c) the positive control — switch the same window to a plain pmacs.buffer.create("*plain*") and require dispatch_idle_for to become true. This is acc16:332-337's half, and it is what rules out the other five clauses in one assertion: a stuck minibuffer, a pending chord, an open menu or an active search would keep the gate false across the buffer switch, so (c) failing is the signal that (b) passed for the wrong reason.

   Bite: delete the set_round_trip_input call in listview.lua:106 / dired.lua:516 and criteria 1–5 all still pass; only this fails, at (b). Falsify (a) by displaying the panel in a side window: (b) then passes with the round-trip mark deleted, which is the whole of P2-4. Falsify (c) by leaving a minibuffer open in the fixture: (b) passes and (c) fails. A daemon-side refusal does nothing for a replica's own mirror, which is why this is pinned through dispatch_idle_for and not through read_only.

   Note what this criterion is NOT. dired_acceptance.rs:999's status(&s).contains("read-only") passes both before and after adoption, because BufferError::ReadOnly and the intercept's own message both contain that substring — the trap docs/dired-stage2-framing.md §3.1 hands to this lane. Criterion 5 is where the distinction is asserted, on the full message text; this one must not be counted as coverage of the adoption.

7. [mutation] A refresh reaches the window, not just the rope — pinned by painting a shrinking render (many rows → one) and asserting row 1 is empty, for each adopter. Revision 2 corrected this criterion's bite (sweep A). Revision 1 claimed it caught a "partial conversion" that kept a bypass_intercept write beside the primitive; that is wrong — such a conversion raises at the bypass write (§2.4, measured) and never reaches a stale paint. The real bite is the one-line mutation delete the notify_buffer_edit_to_windows call in the set_generated_contents binding (src/lua_bindings/mod.rs:3092), which a reviewer can perform.

8. [main] Cursor clamp (Q#GB6). After a shrinking refresh, pmacs.editor.cursor() <= buf:len() and C-p moves. Bite: measured on ad41cf1 — cursor 29, len 2, C-p leaves it at 29. Fails on main today, including for terminal copy mode. 8b. [main] view_top clamp, on a LONGER buffer (Q#GB6, review P2-4). With a window scrolled so view_top sits on line 5, replace "a\nb\nc\nd\ne\nf\n" (12 bytes, 7 lines) with a single line longer than 12 bytes, then require view_top < TextView::line_count(). Bite: a clamp gated on "the buffer shrank" passes 8 and fails 8b, which is the whole of P2-4. Unlike 8, this case is argued from the types and from rebuild_views_for's existing clamp (src/editor_core.rs:1865-1882), not measured — staging it needs a scrolled window.

   8c. [main] Selection-anchor clamp-or-clear, in BOTH clamp sites. New in revision 6 (round 5, P1-5). This is the only one of the three coordinates that PANICS, and #191 reproduced it. Select bytes 0..30 in a generated buffer, have the owner refresh it to two bytes, then invoke copy (region_bytes). Require: no panic; and either a valid region within the new extent, or no selection at all where the clamp collapsed it, per Q#GB6 rule 3.

   Assert it twice, once per site, because they have different callers and neither subsumes the other: through notify_buffer_edit (the generated-write path, which clamps nothing today) and through rebuild_views_for (the *help* / *buffer-list* rewrite path, which clamps cursor and view_top and not selection).

   Bite: measured by #191 — panic at src/rope.rs:145, debug_assert!(end <= self.len()), reached from region_bytes (src/editor_core.rs:4184-4191) via Window::region (src/window.rs:472-479), which clamps neither endpoint. Falsify the clear half by clamping only: the selection then survives as a zero-width region that the user never asked for, which is what src/terminal/view.rs:715-721's collapsed_by_clamp exists to prevent for the terminal's own selection type. Assert the produced region, not merely that the call returned — "did not panic" is satisfied by clearing the selection unconditionally, which would be a different bug.

9. [main] A foreign buffer named *references* is never adopted (Q#GB13). Create a plain buffer of that name with user text, then open the references panel. Assert both halves: the user's bytes survive and an ordinary edit to the user's buffer still lands; and the panel appears under a disambiguated name. Bite: measured — "my precious notes" → "H\nr1", one buffer not two, and the user's buffer is left permanently un-editable. The second half is what fails if adoption is merely made "safe" by skipping the render.

10. [fix-shape] The disambiguation limit raises rather than adopting, matching dired.lua:493-503 / terminal.lua:309-315. Bite: an implementation that falls back to adoption once the limit is exhausted passes 9 and fails this.

11. [main] A disambiguated listview panel still answers RET, g and q (Q#GB18, review P1-2). Continue criterion 9's fixture: with a foreign *references* in place, open the references panel — it appears as *references*<2> — then, through dispatch_key, press g, RET and q in turn and assert the content produced by each: g re-renders (the on_refresh rows appear), RET fires on_visit (assert the visited item, via a probe that records it), and q restores the previous buffer. Bite: this is the criterion that fails against Q#GB13 landed without Q#GB18 — disambiguation alone leaves panel_for_current_buffer looking up panels["*references*<2>"], which was stored under "*references*", so all three commands return early and do nothing. Assert what each command produced, not that it did not raise: every one of the three fails silently, so a "no error" assertion passes against the bug.

12. [main] The q-target capture is not inverted (Q#GB18). The fourth consumer, and it needs its own criterion because it fails open rather than closed. With a disambiguated panel focused, open a second panel (*outline*) and assert the second panel's q returns to the buffer that was current before the first panel, not to the first panel. Bite: with panel_for_current_buffer unable to recognise *references*<2>, listview.open's guard at :118-123 reads "the current buffer is not a panel" and captures the panel as p.prev — the chained-panel q loop the guard's own comment says it exists to prevent. Criterion 11 passes with this bug live, because each command works in isolation; only the two-panel sequence shows it.

13. Q#GB16's fold change — SPLIT into two criteria in revision 4 (review round 3, P2-5), because revision 3's single criterion contradicted its own classification. Revision 3 labelled it [fix-shape] and then wrote that the first half "passes on main ... a dired buffer ... folds fine". If folding succeeds on main, an assertion that it returns false fails on main — which is a main pre-image, the opposite label. The two halves have different pre-images and belong apart.

    13a. [main] A locked generated buffer is not foldable. On a Stage 1 dired listing, pmacs.fold.fold(buf, range) returns false. Bite: this fails on main, where the same call returns true — verifiable by revert, and scripts/bite reports it. It is a regression pin for the intended change: the point is to make sweep C's silent behaviour change into a stated one, so the criterion asserts the new behaviour and the bite is that the old behaviour is different.

    13b. [mutation] …and the refusal says why. The status after 13a's call names the read-only lock, not not a document buffer. Bite: the falsifying mutation is reverting fold.rs:68's string; that is the shape that ships if 13a is written alone — correct behaviour, false explanation. It cannot share 13a's main pre-image because on main the call succeeds and sets no status at all.

14. [structural] No bypass_intercept write remains in dired.lua or listview.lua; listview.ensure_panel contains no find-by-name adoption; and no panels[ subscript remains keyed by a name derived from describe.buffer — the Q#GB18 half. Rides alongside 1–13, never instead: a structural comparison of two authorities does not catch a misrouted consumer, which is why 11 and 12 assert through dispatch_key.

Moved out of Stage 1 in revision 3 and removed in revision 5: the unlock criterion. Revision 2's Stage 1 criterion 11 pinned unlock_generated; revision 3 moved it to Stage 2, revision 4's identity exclusion failed to bound it by generated-lock provenance, and Q#GB7 now ships no binding in either stage.

Stage 2

1. [main] M-x buffer.undo cannot destroy *compilation* / *shell-command* / *search-results* content — and the criterion must assert the exit marker survives, not that the buffer is non-empty. Bite, and this is the whole point: measured, the result of M-x buffer.undo on *shell-command* is [shell exited with code 0] replaced by [output desynced by external edit]. The buffer is still non-empty, so a "not empty" assertion passes with the bug live. The revision guard marks the corruption; it does not prevent it.

2. [fix-shape] A streaming run's incremental writes still land, including CR overwrite semantics (a progress-bar fixture) and erase-to-eol. Assert the produced content, not the absence of an error. Bite: the tempting half-conversion — reset via set_generated_contents, stream via bypass_intercept — raises is read-only at the first append (§2.4, measured).

3. [main] The rope is locked BETWEEN batches, not only after the run. Mid-run, after one output batch has landed and before the next, a bypass_intercept write must be refused and Buffer::is_read_only() must be true. Rewritten in revision 2 (review P1-1). Revision 1 said "attempt an ordinary edit and require the refusal", which passes on main — the intercept refuses ordinary edits today whether or not the rope is locked. A bypass write is the discriminator: it lands on main (compile.lua performs nine of them) and raises once the rope is locked. Bite: a scope-shaped implementation that unlocks for a whole run passes 1 and 2 and fails this. A state predicate, not a geometric readout.

4. [main, and also fix-shape] History is discarded per generated write, asserted past the lock. In a Rust acceptance test, after N batches: buffer.set_read_only(false), then assert buffer.undo() is Err(BufferError::NothingToUndo) and — under --features crdt — that the CRDT reports can_undo() == false; restore the lock. Rewritten in revision 2 (review P1-1). Buffer::undo calls ensure_writable() first (src/buffer.rs:1302) and returns ReadOnly before it ever looks at the stacks, so revision 1's "buf:undo() returns false" passes against an implementation that locks the rope and never clears history. Lifting the lock inside the test is what makes the assertion about history rather than about the lock. tests/terminal_copy_mode_acceptance.rs:582-584 is the existing precedent for a Rust-side lift. This criterion fails on main (where history accumulates) and against the locks-but-never-clears implementation, which is the strongest pairing available.

5. [main] ensure_slot leaves *compilation* locked before any run (Q#GB5). Create the slot without running anything, then require a bypass_intercept write to be refused and is_read_only() to be true. Rewritten in revision 2 (review P1-1): revision 1's "attempt an ordinary edit; without the explicit lock it lands" passes on main, because ensure_slot installs the erroring intercept at compile.lua:266 at creation time.

6. [main] A generated write on a path-backed buffer is refused — on ALL FOUR surfaces, not just the legacy wrapper (Q#GB10; rewritten in revision 3 for review P2-5). Open a file, then, against its buffer, exercise each of:

   surface	call
   the wrapper	pmacs.buffer.set_generated_contents(b, "x")
   insert	b:insert(0, "x", { generated = true })
   delete	b:delete(0, 1, { generated = true })
   replace	b:replace(0, b:len(), "x", { generated = true })

   Each must error, and after each the buffer's contents, is_read_only() and is_modified must all be unchanged (§3.4 exit 1: nothing is touched). Second half, once: after an ordinary edit, autosave still queues that buffer.

   Bite, and this is exactly P2-5's point: a guard placed on Buffer::set_generated_contents rather than on Buffer::apply_generated_edit passes the wrapper row and fails the other three, while the newly public surface could still replace a file buffer's contents, lock its rope and — with mark_clean — make autosave.rs:363 skip it, so a crash loses the user's edits with no recovery slot. Reverting the guard to the wrapper is the one-line mutation that falsifies this. Assert the autosave queue, not just the flag: asserting a value was stored is not asserting anything reads it.

7. [main] mark_clean (Q#GB10). pmacs.describe.buffer(b).modified is false after a generated write on a pathless buffer. Bite: measures true on ad41cf1.

8. [main] Foreign buffers named *compilation*, *shell-command* and *search-results* are never adopted (Q#GB13) — same two-halved shape as Stage 1 criterion 9, plus the limit criterion of 10. Bite: measured on ad41cf1 for *compilation*.

9. [main] A FAILED pmacs.compile.run leaves a foreign *compilation* untouched AND editable. Call it with display = "bogus" against a pre-existing foreign buffer of that name. Bite: measured — today the call raises at compile.lua:754, the contents survive, and the user's buffer is nonetheless left permanently un-editable (ensure_slot ran first and installed an intercept it discarded the handle for). With Q#GB5's lock placed naively it would additionally be emptied and locked. This is the criterion that pins the ordering constraint, and it fails on main today for the intercept half alone.

10. Coverage, not a criterion: both configurations — default and --features crdt — for criteria 1–5 and for the §3.4 transaction criteria 15, 15a, 16 (first half), 16b and 18. CI never enables the feature, so CRDT must not be the only home of any of them.

    Three are irreducibly crdt-only and must say so rather than be quietly written once (revision 4): criterion 16's second half (the Rejected refusal needs a mid-codepoint position, which only loro rejects), 16c (Diverged has no default-feature analogue at all), and Stage 2 criterion 4's can_undo() == false half. Each is paired with a default-configuration sibling that exercises the same cleanup arm through a non-CRDT failure — criterion 15's FailingView is the default-feature route into AppliedThenFailed, and 16b's empty write is the default-feature route into NoOp — so no cleanup arm of §3.4 is reachable only under crdt except Diverged, which by construction cannot be. That pairing is the point of the rule; naming which criteria are single-configuration is what keeps it checkable.

11. [structural] Zero bypass_intercept writes remain in compile.lua and in default.lua's search panel (comments excepted; §1.1's arithmetic is the reference), and neither ensure_slot nor ensure_search_panel contains a find-by-name adoption.

12. [fix-shape] The three intruder tests still assert what they were written to assert after being converted to a Rust-side read_only lift (Q#GB12), rather than being deleted or weakened. Bite: a conversion that drops the intruder edit entirely leaves the desync machinery unpinned while the suite stays green.

New in revision 3 — the transaction and the identity guard. Revision 5 removes the attempted bounded unlock.

13. [structural] No Lua unlock surface is added (Q#GB7). There is no pmacs.buffer.unlock_generated, no exposed set_read_only, and no binding that clears read_only without performing an owner-authorized generated write. Bite: adding the revision 4 binding fails this structural assertion even if it refuses terminal identity buffers; identity_protected == false is not proof that the generated primitive installed the lock.
14. [main] A generated write REFUSES a terminal identity buffer (Q#GB15; sweep C item 1). Open a real terminal; call pmacs.buffer.set_generated_contents(term_buf, "junk") and each of the three { generated = true } mutators against it. Every one must error, and the terminal's contents must be byte-identical afterwards. Bite: this fails on main today — shipped set_generated_contents does self.read_only = false unconditionally (src/buffer.rs:546), overwrites the buffer and re-locks it, and nothing in the tree refuses it. It is the pin for a hole that predates this arc, which is why it is a main pre-image rather than a mutation bite. Falsify by deleting §3.4's exit 2. Criteria 15-16 were VACUOUS in revision 3 and are rewritten. Review round 3, P1-2, is right and the diagnosis is worth stating because it is the second time this arc has shipped a criterion that passes with the bug restored. Both used an out-of-bounds op. Bounds validation is §3.4 exit 4 — before editing_in_progress is set and before the internal lift — so the operation never enters the transaction the criteria claim to test, and an implementation that omits both the flag clear and the error-path relock passes both. Worse, revision 3 argued in the same document that pre-validation makes an invalid range cost nothing; the design decision and the test strategy contradicted each other in adjacent sections. The rule this yields: a criterion must name the exit it drives the implementation to, and that exit must be inside the mechanism under test. Every criterion below names its exit.

Pre-image relabels, revision 6 (round 5, P2-6). Six criteria carried [main] labels that describe what the FIX changes rather than what the BASE does — the classification error this arc has now made twice. The test in every case is: run this criterion's assertions against 300cbc4; do they fail?

criterion	what the base actually does	was	now
15	{ generated = true } is an unknown option key on main — parse_bypass_intercept reads only bypass_intercept, so the call is an ordinary managed edit, and run_managed_edit clears the flag before phase 3. The follow-up edit lands. Passes.	[main]	[mutation]
16	same: no generated path exists to relock or not relock	[main]	[mutation]
16b	an empty write on main is an ordinary managed no-op; the assertions about lock/clean/history describe the fix	[main]	[mutation]
17	invalid-range rejection already preserves history on main, because no clear_history runs on that path at all. Passes.	[main]	[mutation]
18	begin_edit already rejects same-buffer re-entry (src/buffer.rs:726-731). Passes.	[main]	[fix-shape]
21	search already consults pmacs.compile through the optional triple guard (default.lua:991-994). Passes.	[main]	[fix-shape]

None of these criteria is weakened by the relabel — each already stated its falsifying mutation, and the mutation is what gives it bite. What changes is that the document no longer claims a main failure it does not have, which is what scripts/bite would have contradicted. The rule, restated for the third time and now applied by construction: a criterion's pre-image is a fact about the base, established by running it there — not an inference from what the fix is for.

15. [mutation] editing_in_progress is cleared on a failure that ENTERS the transaction (Q#GB17; §3.4 AppliedThenFailed). Attach a Rust-side FailingView — pmacs::view::View is pub (src/view.rs:221, src/lib.rs:141) and Buffer::attach_view is pub, so a test crate can implement one whose on_edit returns Err(BufferError::Intercepted { .. }) — then perform a valid generated write. It fails at the broadcast, after the rope swap. Then lift read_only Rust-side and require an ordinary edit on the same buffer to report the FailingView's message, not is already being edited. Bite: an implementation that returns from the match without reaching §3.4's line (6) leaves the flag set, and begin_edit (:726-731) and apply_edit (:774-779) then refuse every later writable edit to that buffer for the rest of the session. Without the lift, ensure_writable runs before the flag check and both the correct and broken implementations report ReadOnly, which was revision 4's non-discriminating form. Falsify by moving the flag clear inside the Applied | NoOp arm. Assert the next edit's outcome, not the failing call's — the failing call reports the same error either way, which is the whole reason this criterion is about the buffer's state afterwards. Restore read_only after the probe so criterion 16 begins from the specified post-failure state. 15a. [mutation; both configurations] AppliedThenFailed fans out its carried Edit before surfacing the error (Q#GB17; review round 4, P1-1). Display the test buffer in a real window, attach the same FailingView, and issue a valid shrinking generated replace through the Lua binding. The call returns the view's error, but painting the window must show the new shorter contents with no stale-line panic. Under --features crdt, attach a replica mirror before the call and require it to import the owner's crdt_op despite the Lua error. Bite: deleting only the AppliedThenFailed arm's notify_buffer_edit_to_windows call leaves criteria 15 and 16 green — buffer cleanup is correct — while the window keeps stale line ranges and the replica never receives an operation that already changed the authoritative CRDT. This criterion fails that mutation in both directions.

    Revision 6 (round 5, P1-4): the fixture must place a RECORDING view AFTER FailingView in attach order. As written, 15a observes only the window and replica consumers, which are fanned out after the borrow drops — so it cannot see the buffer-attached views that on_edit's stop-at-first-error loop skips. Attach order is FailingView, then a RecordingView whose on_edit appends the Edit it received; assert the recorder saw the edit. Bite for this half: on the base, and under any implementation that keeps view.on_edit(self, &edit)?, the recorder's log is empty while every other assertion in 15a passes — which is precisely why the review could find this gap after 15a had been written and reviewed. Falsify by restoring the ?.

16. [mutation] A generated write relocks on that same failure, and does NOT lock on a refusal (Q#GB17). Two halves, because §3.4 gives them opposite answers and revision 3 gave them the same one.

    • Relock on AppliedThenFailed: after criterion 15's failing write, Buffer::is_read_only() is true and a bypass_intercept write raises. Bite: an implementation that unlocks and returns without relocking leaves the buffer writable, and every criterion about undo silently stops applying. Falsify by deleting self.read_only = true from the AppliedThenFailed arm.
    • No lock on Rejected: on a fresh, writable pathless buffer under --features crdt, a mid-codepoint generated insert is refused; afterwards is_read_only() must still be false and an ordinary edit must land. Bite: revision 3's own design fails this — it relocked unconditionally, so a caller got an error and a locked buffer it never wrote. Falsify by replacing self.read_only = entry_read_only with = true. This half is crdt-only, so criterion 10's coverage rule names it explicitly. 16b. [mutation] A successful no-op still discharges the invariant (Q#GB17; §3.4 NoOp). New in revision 4 (P1-1 direction A). On a pathless buffer, insert text and delete it back to empty so the rope is empty and the undo stack is not; then call set_generated_contents(b, ""). Afterwards: is_read_only() is true, describe.buffer(b).modified is false, and — after a Rust-side lift — buffer.undo() returns Err(NothingToUndo). Bite: revision 3's predicate fails all three. The no-op arm (src/buffer.rs:1245-1253) returns Ok without bumping revision, so revision 3 skipped the clear and the mark_clean and left a locked, modified buffer with poppable history. Falsify by restoring if self.revision() != rev_before. This is not a hypothetical path: Q#GB5 prescribes exactly this call in ensure_slot. 16c. [fault-injection, crdt-only] A CRDT mid-transaction failure is distinguishable and contained (Q#GB17; §3.4 Diverged). New in revision 4, made stageable and mandatory in revision 5. In a src/buffer.rs unit test, drive the private delete→insert classifier with a delete closure that succeeds and an insert closure that returns the same error shape as loro. Require Diverged, then drive that outcome through generated cleanup and require: a distinct error variant, not CrdtRejected; the buffer left read_only; and undo history not cleared. Bite: folding the variant into Rejected restores a fresh buffer's entry_read_only = false; the test must assert the lock post-state so that mutation fails, not only the error discriminant. cargo test --lib --features crdt is the explicit gate. No public fault-injection API is added, and there is no four-variant fallback.

    Revision 6 adds two halves (round 5, P1-2 and P1-3), and the seam changes shape to admit the first.

    • The seam is no longer a "delete→insert classifier". Per §3.4 it is the crdt_mutated flag carried out of apply_to_crdt_then_normalize_bytes, so the injectable closures are the CRDT primitives and export_updates_since. New half: force export_updates_since to fail after a successful crdt.delete and require Diverged. Bite: revision 5's classifier returns Rejected here, which restores a fresh buffer to writable while the CRDT and rope disagree — the exact harm round 4's P1-3 withdrew the four-variant fallback to prevent. This half fails against revision 5 as written, not merely against a hypothetical implementation.
    • New half: the replica path is quarantined. After forcing Diverged, require export_buffer_snapshot(&editor, buffer_id) (src/daemon.rs:2693-2708) to return None, and require the initial-attach loop to skip the buffer. Bite: revision 5 fails this — it asserted read_only was containment, and neither export path reads read_only, so a replica attaching after the divergence received the divergent CRDT as authoritative while every daemon window still painted the old rope. Falsify by deleting the crdt_quarantined guard from either export site. Assert the absence of a published snapshot, not the presence of the flag — asserting a value was stored is not asserting anything reads it.

17. [mutation] An invalid-range generated write does NOT destroy undo history (Q#GB17). On a pathless buffer with two ordinary edits already on the stack, call b:delete(0, b:len() + 1000, { generated = true }); require the error, then lift read_only Rust-side and require buffer.undo() to succeed and restore the prior contents. Bite: this fails against the shipped ordering transplanted verbatim — set_generated_contents calls clear_history() unconditionally (src/buffer.rs:551-553), so a call that changed nothing would wipe the user's history. It is the concrete cost review P1-1 asks the ordering to state, and pre- validation (§3.4 exit 4) is what pays it.

18. [fix-shape] A re-entrant generated write is refused (Q#GB17). From inside an add_intercept body on buffer X, call X:insert(0, "x", { generated = true }); require ConcurrentEdit, and require the outer edit to complete normally afterwards. Bite: omit the gate and the inner write mutates the rope while run_managed_edit phase 3 is holding an op computed against the pre-edit InterceptContext (src/lua_bindings/mod.rs:1477-1487); the visible symptom is the outer edit landing at the wrong offset, so assert the resulting text, not the error.

19. [main] is_generated_buffer recognises a disambiguated buffer (Q#GB18). With a foreign *compilation* in place, run M-x compile; the run lands in *compilation*<2>. From inside that buffer, run M-x compile again (the g-recompile path) and require the q target still to be the user's original buffer — not *compilation*<2>. Bite: with is_generated_buffer still comparing names, compile.lua:762's guard reads the disambiguated buffer as "not generated" and re-captures it, so q returns the user into a compilation buffer. Assert where q lands, not whether the predicate returned a boolean.

20. [structural] Zero name comparisons remain in the Class 1 sites. pmacs.compile.is_generated_buffer contains no d.name ==, and listview.lua contains no panels[d.name]. Rides alongside 11–19, never instead.

21. [mutation] A REFUSED generated write leaves the fold closed (Q#GB3; round 5, P2-7). New in revision 6. Fold a region of a file-backed buffer, then, from inside an interactive command so InteractiveCommandOrigin::current() is Some, attempt b:replace(s, e, "x", { generated = true }) on it. Q#GB10 refuses it. Require: the error, and #pmacs.fold.folds(b) == 1 with the same range still stored. Bite: revision 5's ordering fails this. It called unfold_before_interactive_lua_edit before run_generated_edit, so the fold opened and then the write was refused — a visible side effect from an operation whose contract says "before any state change". Falsify by moving generated_preflight back below the unfold. Assert the fold count and range, not just that the call errored: the error is identical either way, which is exactly why this needs a state assertion rather than an outcome assertion.

    Repeat once with the identity-protected refusal (a terminal identity buffer) so the criterion pins the preflight rather than Q#GB10's path check specifically — a preflight that hoists only the file_path test passes the first half and fails this one.

22. [fix-shape] Search works with no pmacs.compile present (Q#GB18; review round 3, P2-4). New in revision 4. Build a LuaHost directly, call attach_editor, and — with compile.lua never loaded — run pmacs.project.search. It must not raise, and the q target must be the buffer that was current before the search. Then assert the harness premise explicitly: pmacs.compile == nil, so a fixture that later gains the runtime sequence fails loudly rather than passing as an ordinary editor test. Bite: revision 3's design fails this at the first search — ensure_search_panel called pmacs.compile._register_generated_buffer unguarded, and pmacs.compile is nil here. Falsify by replacing the symmetric guard with a direct call. This configuration is reachable with the pattern nine existing test files already use (LuaHost is pub; default.lua loads at src/lua.rs:250-251, compile.lua at src/editor.rs:704), and no ordinary acceptance test reaches it, because every EditorState::new loads both — which is exactly why the guard it defends was invisible enough for revision 3 to remove.

---

7. Bets

• That a per-op clear_history is not a throughput problem. Argued from create_undo_manager being UndoManager::new(doc) (O(1) in document size) and from at most one v0.1 entry existing per clear. Measured in Stage 2, not asserted here.
• That converting compile's nine sites does not disturb its byte anchors. The conversion changes authority, not op shape, position or count — emit_text's slot.out_pos arithmetic is untouched. The bet is that nothing else in the module reads read_only indirectly; criteria 2 and 3 are what test it.
• That { generated = true } sitting beside { bypass_intercept = true } is clearer than replacing it. Q#GB2.
• That duplicating the rope's bounds arithmetic in validate_op_bounds is worth what it buys (§3.4). The alternative to a pre-check is a dry run, and Rope offers none; the cost is one O(1) helper that must stay in step with Rope::insert / delete / replace's own bounds rules (src/rope.rs:174-230). What it buys is criterion 17 — an invalid range that costs no history. If the user prefers no duplication, the fallback is to accept the shipped unconditional clear and drop criterion 17, which should be a stated trade rather than a silent one.
• That one extra bool on Buffer is the right size for terminal identity protection (Q#GB15), rather than the enum §2.11's three policies really want. It is deliberately not lock provenance and enables no unlock. The bet is that the enum is a separable refactor; if it is not, the field becomes churn the refactor has to undo.

8. Deferred (named)

• The REPL's undo exposure (Q#GB8), with the §2.5 measurement.
• Class C: *buffer-list*, *help*, *workers* (Q#GB9), and specifically *help*'s two independent owners (§1.4) — a Rust writer at src/help.rs:354 and a Lua writer at default.lua:1239, disagreeing on mark_clean, each with its own copy of the name constant. Whoever takes Class C decides who owns *help* before they decide what it writes with.
• The CRDT Replace mid-transaction divergence (§3.4's Diverged). EditOp::Replace is two loro ops — crdt.delete then crdt.insert (src/buffer.rs:1140-1163) — and if the first succeeds and the second fails, "the CRDT is mid-transaction ... and the rope is unchanged. This is an invariant violation", in the code's own words. This arc names the state and contains it; it does not repair it. Repair means either a single transactional splice or a reconciliation pass, both loro-level work with no bearing on generated buffers specifically. It reaches apply_edit and apply_edit_skip_intercepts today and is reported as an ordinary CrdtRejected, so nothing in the tree currently distinguishes it — which is the smaller half this arc does fix. Named here so the next lane starts from the citation rather than the symptom.
• Suppress-rather-than-clear history recording, if Stage 2's measurement says the per-op clear costs anything.
• read_only in describe.buffer (Q#GB14) — separable, no new capability, not required by this arc.
• Replacing read_only: bool with a policy/provenance enum (Q#GB15's rejected alternative). It is the shape §2.11's three policies actually want, and it would let document_bytes ask the question it means instead of the question the flag happens to answer (Q#GB16). Rejected here as a refactor this arc would be smuggling; named as the right eventual shape.
• Ordinary edit fan-out after a post-apply View::on_edit error. The richer internal outcome makes the pre-existing loss explicit, but apply_edit and apply_edit_skip_intercepts keep their public Result<Edit, BufferError> surface in this arc. Generated bindings retain and fan out the edit; generalizing that contract across every Rust and Lua edit caller is separate work.
• The five *scratch* find-or-create copies (§2.10 Class 5) — default.lua:581, :1145, listview.lua:190, compile.lua:1052, dired.lua:914. Correct only while *scratch* stays unowned and undisambiguated; nothing in the tree connects them, so a future lane that gives *scratch* an owner breaks all five at once.
• *scratch* can be permanently locked by any Lua caller today. Sweep B (revision 2) found this and revision 2 treated it as a reason to ship an unlock. It is a pre-existing exposure: pmacs.buffer.set_generated_contents is already public on main and already locks any buffer id it is handed. This arc neither creates it nor closes it, and revision 5 deliberately adds no unlock because identity_protected cannot prove who installed the lock. Recorded as a standing hazard rather than as this arc's work, which is the correction revision 3 began and revision 5 completes.
• docs/agent-handoff.md §4's inventory is keyed by bypass_intercept and therefore misses Class C, and its headline "four writer mechanisms" is five once src/help.rs:354 is counted (§1.4). Not this lane's file to edit mid-flight; carried in the PR body.
• Removed from this list in revision 3: COHERENCE.md §14's listview consumer list. PR #189 landed the correction (§1.5). A merged correction is removed, not relabelled.
• Wdired's generated→editable transition. Revision 1 deferred it, revision 2 made a general unlock Stage 1 work, revision 3 moved that capability to Stage 2, and revision 5 removes it. The wdired consumer stays deferred to dired Stage 3, which must frame an owner-specific transition or the eventual policy enum rather than inheriting a Lua clear of read_only.

9. Coherence impact (COHERENCE.md §20)

Section served: §14 Coherent Workbench Primitives, and specifically its Output channel bullet, which is where this caveat is already recorded ("four writer mechanisms have not yet adopted it and remain emptiable"). This arc discharges that entry for Class A and replaces its "a streaming variant of the primitive that does not exist yet" with one that does. §14's list-primitive bullet is touched too: listview is called "the strongest coherence asset in the UI layer", and it is currently emptiable by one keystroke.

• Priority 5 (finish the workbench convergence) is the priority this serves. It is a correctness debt inside an existing primitive rather than a new primitive, so it is wiring, not model.
• §14 consistency, added in revision 2: Q#GB13 makes three writers honour an ownership rule the tree already states (terminal.lua:300-305) and already implements twice (dired, terminal). That is §14's thesis applied to a discipline rather than a view — five generated-buffer owners converging on one identity rule instead of three of them inventing find-by-name.
• §6 interaction islands — none added. No new keymap scope, no new dispatch shadow, no new precedence rung. The count stays at six. This arc deliberately does not add undo-chord rebindings anywhere; the measured point of the bug is that rebinding chords was never the fix.
• §11 configuration registry — no new settings, no adoption change.
• §2 golden journey — step 6 (compile) and the dired/browse steps are touched only in the sense that their buffers stop being destructible. No journey step opens or closes.
• Background-work attribution — unchanged. compile.lua's process pump and the grep stream keep their existing ownership.
• Protocol — no change. Nothing new crosses the wire; the fan-out reuses queue_daemon_origin_crdt_op.
• §14 correction — LANDED, not owed. PR #189 (main @ 7586905) corrected the listview consumer list and moved the scorecard row from ✓ to ◐ (§1.5). What remains owed on merge is the handoff §4 table, whose four-row inventory is keyed by bypass_intercept, therefore misses Class C, and undercounts the mechanisms by one (src/help.rs, §1.4).

9b. Cross-lane boundaries

Four lanes touch adjacent ground. The boundaries below are settled elsewhere and are recorded verbatim rather than re-decided here.

#188 → #191 — the acceptance-contract boundary. Added in revision 6:

#188 (framing) owns the acceptance contract. #191 (Stage 1 implementation) may not restate, narrow, or reclassify a criterion — it adopts what the framing says. Where an implementation found a criterion impossible, the framing is revised and re-approved first; the implementation then follows. The selection-anchor clamp is Q#GB6's to specify and #191's to implement, and both must describe the same rule.

What that boundary has already had to settle, twice, in this revision. Both are cases where #191 was right about the tree and wrong about who decides:

1. Stage 1 criterion 5 was impossible (round 5, P1-1). #191 reached that independently and restated the criterion locally — which is exactly what the boundary forbids, and which left #191's tests and ledger describing a contract this document did not carry. Criterion 5 is settled here, in revision 6; #191 adopts the (a)/(b) split rather than its local restatement. The same applies to the criterion 7 divergence #191's review found.
2. The selection anchor (round 5, P1-5). #191 reproduced the panic; Q#GB6 in this revision is the specification, including the clamp-or-clear rule and the two clamp sites. #191 implements that text. If implementation finds it wrong, the correction lands here first.

Why this boundary is worth its overhead. A framing and its implementation disagreeing about a criterion is not caught by either lane's gates — both are green against their own description. It is caught only by a reviewer holding the two documents side by side, which is what happened, and only after the divergence had reached #191's tests and its ledger entry.

#186 / #171 — recorded, not this lane's:

#186 owns the urgent pre-filesystem refusal for synchronous apply_resource_op. #171 later owns full post-delete lifecycle reconciliation, including the async race where a buffer becomes modified after dired dispatch.

#171 → #188: Q#DR25 is deferred INTO this lane, and revision 3 is the first revision to say so. Revisions 1 and 2 of this document never mentioned Q#DR25, #171, or dired Stage 2 at all — a gap, since the other lane had already handed the work over. Read against #171 revision 7 (fd7ae37, pushed 2026-07-28), which is that document's current state:

• #171 §3.1 and its Q#DR25 entry state that dired's listing becoming a genuinely immutable generated buffer is "not Stage 2's decision to make", that it is "owned by the generated-buffer-immutability lane", and that "Stage 2 does not implement it, does not gate on it, and carries no acceptance for it." This document's Stage 1 claims exactly that work (dired.lua:369-372 adopting set_generated_contents), so the claim is live and the two documents agree.
• Neither ordering creates a conflict, per #171 §3.1: Stage 2b changes paint's callers, this lane changes paint itself. If this lane lands first, Stage 2b rebases onto a paint that already writes through the primitive; if Stage 2b lands first, this lane adopts a paint with more callers and needs no change to them.
• One inherited fact this lane must not lose, recorded in #171 §3.1 as "a trap for that lane's acceptance": tests/dired_acceptance.rs:969's dired_buffer_is_read_only_and_round_trips_input asserts status(&s).contains("read-only"), and BufferError::ReadOnly renders as buffer `{name}` (id {id:?}) is read-only — so that test passes both before and after the adoption and is not coverage of it. It is cited in Stage 1 criterion 6 as the model for the document-window premise only; the note at the end of that criterion says so explicitly.
• One difference worth flagging, not a conflict. #171 revision 5's withdrawn plan had dired's paint adopt the primitive "dropping the erroring intercept". This document keeps the intercept at both adopters, per the layering terminal.lua:351-366 states. Since rev 6 withdrew the decision from #171 entirely, this lane owns it and there is nothing to reconcile — but a reader who finds rev 5's phrasing should know it was superseded, not contradicted.

10. Verification plan

Full gate suite per CLAUDE.md for each PR separately:

cargo fmt --check
cargo clippy --workspace --all-targets -- -D warnings   # own step
cargo test --lib
cargo test --lib --features crdt
cargo test --test <the touched acceptance suites>
cargo test --test m4_acceptance -- --skip basedpyright
PMACS_REQUIRE_GPU=1 cargo test -p pmacs-gpu
git diff --check

Plus, per stage:

• Stage 1 — cargo test --test dired_acceptance and --test listview_acceptance (the suite exists; tests/listview_acceptance.rs), plus --test terminal_copy_mode_acceptance if Q#GB6 lands, since the clamp changes the shipped snapshot path. Plus --test folding_acceptance, added in revision 3: Q#GB16 touches src/lua_bindings/fold.rs and read_only_buffer_is_rejected (tests/folding_acceptance.rs:570) pins the guard the lock now reaches. Stage 1 touches Rust (Q#GB6's clamp; Q#GB16's status string), so cargo test --lib and --lib --features crdt are load-bearing for it rather than formalities.

• Stage 2 — cargo test --test compile_mode_acceptance and --test compile_mode_crdt_acceptance, plus --test terminal_copy_mode_acceptance (the set_generated_contents reimplementation, mark_clean, and Q#GB15's write-direction refusal all reach it), plus --test vterm_stage1_acceptance — added in revision 3, because tests/vterm_stage1_acceptance.rs:139,175,290 are the shipped assertions about a terminal identity buffer's read_only and Q#GB15 changes what may touch it. The search panel has no suite of its own — grep -rln 'search-results' tests/ returns only compile_mode_acceptance.rs and m4_acceptance.rs, so Stage 2's search-panel criteria need a new home rather than an existing one to extend.

• Run the crdt-gated terminal suite explicitly, as its own step. Added in revision 4 (review round 3, P1-3).

  cargo test --test terminal_copy_mode_acceptance --features crdt
  

  This is not covered by any other line in the gate list. acc16e_a_refresh_queues_the_owners_write_for_replica_mirrors is #[cfg(feature = "crdt")] (tests/terminal_copy_mode_acceptance.rs:567), and it is the shipped consumer of the lift-and-restore idiom that revision 3's generated_lock would have broken. A default-feature run of the same suite never compiles it, so a green sweep proves nothing about the seam P1-3 is about. Judge it by whether the test count includes acc16e, not by the verdict alone.

• Stage 2 additionally needs a crdt run of whatever suite hosts the §3.4 transaction criteria, for criterion 15a's replica half and criterion 16's second half. Criterion 16c lives in cargo test --lib --features crdt because its private delete-success/insert-failure classifier is a unit-test fault seam, not a public acceptance input. Same reasoning, same failure mode.

• Run scripts/bite on every criterion expressible as a test today, and read its new exit codes (main @ 64883eb, PR #192). Exit 0 is a real bite; 1 is vacuous; 3 is NO CONTROL — the named tests did not pass on the working tree, or none ran, which is the filter-matches-nothing case that used to read as success; 4 is INCONCLUSIVE (MIXED). The swapped run is labelled OK (assertion) or OK (COMPILE). Prefer OK (assertion) and treat OK (COMPILE) as a prompt to narrow the swap to a file that builds both ways — a compile break proves the swap changed something, not that the test discriminates. Every "falsify by …" in §6 names a one-file change precisely so it can be run through this rather than asserted.

• Run scripts/bite on every criterion expressible as a test today. Stage 1 criteria 1–3, 8, 9 and Stage 2 criteria 1, 7, 8, 9, 14 have main pre-images and can be falsified by revert; the rest are fix-shape or mutation bites and each names its mutation inline. A criterion whose bite cannot be stated as either is not finished.

• Do not gate any new test on #[cfg(feature = "crdt")] unless it genuinely needs CRDT. CI never enables the feature — measured at ad41cf1, 276 tests are dark as a result:

  cargo test --all-targets --no-default-features --features lua54 -- --list \
    | grep -c ': test$'          # 3251   (CI's exact flags)
  cargo test --all-targets --no-default-features --features lua54,crdt -- --list \
    | grep -c ': test$'          # 3527
  

  3,527 − 3,251 = 276. Revision 3 deliberately does NOT re-quote this at 7586905. The base moved (#189, COHERENCE.md only, which adds no tests), so the reading is very probably unchanged — but "very probably unchanged" is the reasoning the ledger warns against, and a framing doc is not the authority for this number in any case. Treat 276 as a reading taken at ad41cf1, not as a constant. Re-measured in revision 2 (review P2-5). Revision 1 quoted 264, which docs/active-work.md:107-115 labels historical (#168's reading at 1b6a084) and explicitly warns against: "the number moves with every merge and must be re-measured, not quoted." The ledger's own most recent figure is 273 at 74301d1; this arc's base is later, and the number should be re-measured again rather than quoted from here.

• Judge the touched suites by elapsed time as well as verdict where they reach for a sibling binary (docs/agent-handoff.md §5).

• Commit before gating: cargo fmt after a commit splits the worktree from the branch and git diff --check will not catch it.