69 KiB

Raw Blame History

Generated-buffer immutability

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

Revision 2 — scouted against canonical githubsucks/main @ ad41cf1, 2026-07-28. Every claim below about pmacs was executed, not read. The reproductions in §0 and §2 are transcripts of throwaway probes run in this worktree at ad41cf1 and deleted before each commit; the counts in §1 are whole greps with the arithmetic shown, never | head.

Revision history

Revision 2 answers 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:1853-1857`), 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 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).

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:2575) → 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.

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

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.

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.

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:

*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.
*help* — show_help_text (default.lua:1239), same plain delete-all + insert-all (:1245-1246). No intercept.
*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.

1.5 A correction to `COHERENCE.md` §14 (not edited here)

§14 states that "references, outline, buffer-list, and project-search all use" listview. Measured: pmacs.listview.open has three production callers, all in lsp.lua — *references* (:2056), *outline* (:2102), *lsp-help* (:2513). *buffer-list* is hand-rolled in default.lua (render_list, :387) and *search-results* is the independent grep panel. Two of §14's four examples are wrong, and *lsp-help* is missing. COHERENCE.md is not this lane's file to edit; recorded here and in the PR body.

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() (:559), 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:1814) updates each window's TextView and overlays. It does not clamp win.cursor or win.view_top. Only rebuild_views_for (:1843) does, and its doc comment says so explicitly (:1841-1842). 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:1853-1857); 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:1436,1984,2128,2163, src/daemon.rs:2976). 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.

Also found in the same sweep, and reused below: src/instance_buffer.rs:401 ("rendered buffer must be marked clean") is a third generated-buffer writer that already marks clean, alongside workers_buffer::render.

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) -> Result<Edit, BufferError> — 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: lift read_only → apply_edit_skip_intercepts(op) → clear_history() → re-assert read_only → return the Edit. The binding then fans it out through the notify_buffer_edit_to_windows call it already makes (src/lua_bindings/mod.rs:1291, :1302, :1322), after the borrow has dropped.

Buffer::set_generated_contents(bytes) is reimplemented as apply_generated_edit(Replace { range: 0..len, bytes }). It keeps its name, its signature, its doc comment and its tests — 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, immediately, by the binding that already does it. 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. This is exactly today's fan-out count — the conversion changes authority, not cardinality.

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. 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. History clearing goes to loro's UndoManager. Nothing new.

How do the returned edits reach the fan-out without a live registry borrow? By construction, unchanged since #178: 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. apply_generated_edit slots into the same place apply_edit_skip_intercepts occupies now.

4. Decisions

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, signature and tests. 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 — Routing is otherwise identical to bypass. A generated write goes through run_buffer_edit's bypass arm (src/lua_bindings/mod.rs:1368), including unfold_before_interactive_lua_edit. The unfold guard already requires InteractiveCommandOrigin::current() to be Some, which is false for the process.after-tick pump and true for M-x compile; folding a *compilation* buffer is possible, so changing this would be a silent behaviour change to a pinned seam for no reason this lane owns.

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.

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:1853-1857, 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.

Recommended for Stage 1, because Stage 1's adopters refresh shrinking panels constantly and because it fixes terminal copy mode retroactively. 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.

Q#GB7 — Ship a one-way unlock in Stage 1. Revision 2 reverses revision 1's recommendation, and sweep B is why.

The wdired case is unchanged: dired Stage 3 (docs/dired-framing.md §5) makes a dired buffer editable by removing the read-only intercept and swapping the major mode; once dired's rope is read_only that is no longer sufficient, and §1.3 measured that no Lua binding can clear read_only. Revision 1 deferred the binding on the grounds that "a binding whose only consumer does not exist yet cannot be pinned against a real caller".

Sweep B falsifies that reason. The two halves of the protection are not symmetric:

the intercept half is removable — remove_intercept (src/lua_bindings/mod.rs:3433), which the REPL actually calls (repl/init.lua:325-327);
the rope half is one-way from Lua, permanently.

And { generated = true } is public Lua callable on any buffer id. Even with Q#GB10's pathless rule, a third-party package — or a typo in one — can permanently lock *scratch*: pathless, the default buffer, and the quit target of five separate code paths (dired.lua:914, compile.lua:1052, listview.lua:190, default.lua:585, :1150). That is a caller-visible failure that exists today, has a consumer, and is directly pinnable — which is exactly what revision 1 said the binding lacked.

Recommendation: pmacs.buffer.unlock_generated(buf) — one-way, clears read_only and nothing else — shipped in Stage 1. One-way rather than set_read_only(buf, on) because unlocking removes protection and can therefore never brick anything, whereas a settable lock reintroduces precisely the "lock with no door" trap docs/agent-handoff.md §4 warns about. It is strictly weaker than the setter and strictly sufficient for both consumers (the brick escape, and dired Stage 3's mode swap).

If the user prefers the symmetric setter instead, note that pmacs.buffer.set_read_only(buf, on) contradicts docs/agent-handoff.md §4's "there is deliberately no Lua set_read_only" and needs an explicit ruling rather than a quiet addition — though the objection behind that invariant ("it also refuses the owner's refresh") is answered once { generated = true } ships.

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:3704) 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.

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.

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.
Prerequisite, in this PR (Q#GB7): pmacs.buffer.unlock_generated(buf), a one-way clear of read_only. It is the escape from a bricked buffer and dired Stage 3's mode-swap door. Sweep B upgraded this from a deferral.
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.

Revision 2 grew Stage 1 by two prerequisites and one reversal. Both additions are load-bearing for the lock rather than adjacent to it: the ownership rule is what makes locking safe, and the unlock is what makes a mistake survivable. Stage 1 is no longer a pure-Lua change — the unlock is a new binding — and §5's earlier claim that it was has been corrected below.

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 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.
Buffer::apply_generated_edit + the { generated = true } option + set_generated_contents reimplemented over it.
Q#GB10's path-backed refusal and mark_clean — one rule, both halves, since the refusal is what makes the flag change safe.
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:

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.
Q#GB7 (unlock) is now built in Stage 1, not merely recorded — revision 1 had this backwards. Sweep B found the pinnable consumer revision 1 said did not exist.
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 no longer pure Lua (Q#GB7's binding), which revision 1 claimed and revision 2 withdraws.

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

[main] C-/ cannot empty a listview panel. Driven by dispatch_key, not a Lua call. Bite: measured — "H\nrow-one\nrow-two" → "".
[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.
[main] C-/ and M-x buffer.undo cannot empty a dired listing. Bite: measured — one undo takes the listing to "".
[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.
[fix-shape] An ordinary edit is refused by the INTERCEPT, not by the rope — assert on the message text, which distinguishes them. Measured, both forms: the intercept produces intercept rejected the edit: ... listview.lua:102: *probe-panel* is read-only; the rope produces buffer `*probe*` (id BufferId(4)) is read-only. Bite: an adopter that deletes the intercept and relies on the rope passes 1–4 and fails this. The layering at terminal.lua:351-366 requires the named error to survive.
[fix-shape] set_round_trip_input is still set on both. Pinned ungated via dispatch_idle_for reporting false while the panel is focused — the shape tests/terminal_copy_mode_acceptance.rs criterion 16 uses, which needs no CRDT. Bite: delete the set_round_trip_input call and 1–5 all still pass; only this fails. A daemon-side refusal does nothing for a replica's own mirror.
[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.
[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:1853-1857), not measured — staging it needs a scrolled window.
[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.
[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.
[main] The unlock is real and is narrow (Q#GB7). On a plain buffer with no intercept: set_generated_contents locks it (a bypass write raises), unlock_generated releases it (a bypass write lands), and an ordinary edit then lands too. Separately, on a listview panel: after unlock_generated, an ordinary edit is still refused by the intercept, asserted on the message text per criterion 5. Bite: a no-op unlock fails the first half; an unlock that also tears down the intercept — "unprotect" rather than "unlock" — fails the second.
[structural] No bypass_intercept write remains in dired.lua or listview.lua, and listview.ensure_panel contains no find-by-name adoption. Rides alongside 1–11, never instead.

Stage 2

[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.
[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).
[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.
[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.
[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.
[main] A generated write on a path-backed buffer is refused (Q#GB10). Open a file, then call set_generated_contents on its buffer: the call must error and the buffer's contents, lock state and is_modified must all be unchanged. Second half: after an ordinary edit, autosave still queues that buffer. Bite: without the guard the call replaces the file buffer's contents, locks the rope, and — with mark_clean — makes autosave.rs:363 skip it, so a crash loses the user's edits with no recovery slot. Assert the autosave queue, not just the flag: asserting a value was stored is not asserting anything reads it.
[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.
[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*.
[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.
Coverage, not a criterion: both configurations — default and --features crdt — for criteria 1–5. CRDT must not be the only home of any of them; CI never enables the feature.
[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.
[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.

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.

8. Deferred (named)

The REPL's undo exposure (Q#GB8), with the §2.5 measurement.
Class C: *buffer-list*, *help*, *workers* (Q#GB9).
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.
COHERENCE.md §14's listview consumer list is wrong (§1.5), and docs/agent-handoff.md §4's inventory is keyed by bypass_intercept and therefore misses Class C. Neither file is edited here; both carried in the PR body.
Removed from this list in revision 2: wdired's unlock. It is now Stage 1 work (Q#GB7), because sweep B found it a pinnable consumer.

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 owed: the listview consumer list (§1.5) and, on merge, the handoff §4 table, whose four-row inventory is by bypass_intercept and therefore misses Class C.

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, plus --test terminal_copy_mode_acceptance if Q#GB6 lands, since the clamp changes the shipped snapshot path. Stage 1 now touches Rust (Q#GB7's unlock_generated binding and Q#GB6's clamp), 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 and mark_clean both reach 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.
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. 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.

69 KiB Raw Blame History Unescape Escape