pmacs/docs/agent-handoff.md

Agent handoff — cross-machine continuity

Last updated: 2026-07-28, after the M4 config-sink race fix (#174) and bottom-panel Stage 2B-1 (#184) merged; the canonical landed base is 0442d78. #174 is test-only. #184 is the substantive one — the reserved protocol-v21 bottom-panel wire family, dark by construction, with the production handshake deliberately still advertising v20 — following the Journey/GPU directory-target ratchet (#183), following Journey Stage 1a (#182), which made directory startup one coherent local/daemon/GPU path and incorporated the terminal configuration + copy mode landed-doc work (#180); following terminal copy mode (#178) — C-c C-t materializes a terminal's whole retained range into an ordinary buffer, plus Buffer::set_generated_contents, the first genuinely immutable generated-buffer write path — and its landed-doc pair (#168); following Lean 4 Stage 4a (#179) — the typed-edit consumer chain — and bottom-panel Stage 2A (#177), the classified census routing that makes every Projection-class consumer ask primary_document_window; the bottom-panel Stage 2 framing (#175), terminal configuration Stage 1 (#173) — profiles, scrollback, a per-terminal configurable escape key, and the C-c t opening binding — Lean 4 stages 3a and 3b (#167, #170), pmacs' first Lean language server; the GPU terminal input fix (#166), the double terminal-layout sync that made a GPU terminal untypable; the CRDT undo repro (#157), the inline-math landed-doc refresh (#172), the inline-math slice (#158), the first mathematical typesetting in pmacs; dired Stage 1 (#165), Lean 4 Stage 2 (#161), the dired framing pair (#163/#164), find-file (#162) — the dired arc's Stage 0 — COHERENCE.md (#163), Lean 4 Stage 1 (#160), the minimap blank-slab fix (#159), bottom-panel Stage 1 (#155), the inline-math re-scout (#154), the vterm PTY-flake fix (#153), and the GPU initial-target doc refresh (#152); and before that GPU initial-target (#148, protocol v20), following folding Stage 2 (#149) and its landed-doc refresh (#150), web grammars HTML+CSS (#146), the LaTeX Stage 1 / inline-math framing pair (#144/#145), folding Stage 1 (#142), one-command GPU invocation (#141), the documentation refresh (#140), Vterm Stage 3 (#135), tab-width rendering parity (#137), locals-query processing (#134), modeline detection (#132), mode system wiring (#129), config registry (#127), Vterm Stages 1–2 (#126/#130), and completed Themes Arc 4 (#120/#124/#125). This file is the bridge between development machines. If you are an agent reading this on a fresh clone: this document plus the docs/*-framing.md files ARE your memory. Read this fully before taking on work, seed your persistent memory from it, and update this file (and commit it) whenever project state changes materially — the next machine reads it the way you just did.

For volatile branches, checkpoints, verification, and recovery commands, read docs/active-work.md immediately after this file.

1. Where the project stands (2026-07-28)

• main @ 0442d78 (the M4 config-sink race fix #174 — test-only — atop bottom-panel Stage 2B-1 #184, the Journey/GPU directory-target ratchet #183, Journey Stage 1a #182, incorporating terminal configuration + copy mode landed docs #180, Lean 4 Stage 4b #181, the dired Stage 1 landed docs #169 and the PTY-terminate diagnostic #176, terminal copy mode #178, the GPU-terminal-input landed docs #168, Lean 4 Stage 4a #179, bottom-panel Stage 2A #177, the bottom-panel Stage 2 framing #175, terminal configuration Stage 1 #173, Lean 4 Stage 3b #170, Stage 3a #167, the CRDT undo repro #157, the inline-math landed-doc refresh #172, the bottom-panel landed-doc refresh #156, the inline-math slice #158, dired Stage 1 #165, the GPU terminal input fix #166, Lean 4 Stage 2 #161, the dired framing #164, COHERENCE.md #163, find-file #162, Lean 4 Stage 1 #160, minimap blank-slab #159, bottom-panel Stage 1 #155). Protocol schema support is now v6..=v21, while the production server-first Hello still advertises v20. Those are two different facts and #184 landed only the first: the v21 bottom-panel wire family exists, is gated in both directions, and has no producer, no consumer, and no capability behind it. Compatible v21 activation and the production advertisement move belong to Stage 2B-3. The bullets below describe the arcs in their own terms; this line is the head-of-main anchor.

• COHERENCE.md is now required reading and a required framing input — #163. It carries the product-coherence thesis, an audited scorecard, per-concern gaps, and §20's priority order, and it is the standard new work is evaluated against. Per CLAUDE.md, every new framing doc must state its coherence impact — journey steps touched, interaction islands added, config-registry adoption, background-work attribution. Its §2 grades the golden journey; Journey Stage 1a moved that grade off "broken at step 3" — see the arc bullet below.

• Journey arc (P1) — Stage 1a LANDED (docs/journey-stage1a-framing.md). pmacs . opens a directory instead of exiting 1, on one path: resolve_target_buffer gained a ResolvedTarget::Directory arm ahead of the load, EditorState::open became a caller of it rather than a parallel implementation, and the daemon/GPU bootstrap shares the same arm. Which surface handles a directory is the path.open-directory chain with dired as a replaceable fallback slot. tests/journey_acceptance.rs is the new cross-subsystem ratchet (steps 2, 3, 5 seeded; stages add rows, none removes them). No protocol change.

  • A hook a builtin subscribes to can never be first-claimant-wins for users. HookRegistry::add only appends and builtins load before init.lua, so a dired subscription would always claim before any user listener. That is why dired is a slot (pmacs.path.directory_handler) and not a subscriber — and why clearing the slot has to leave startup succeeding with a status, not exiting 1.
  • A raise and a false are indistinguishable in proceed. run_short_circuit returns proceed = false for both; only HookOutcome.errors separates them, and it decides whether to report, not whether to fall back. Getting this backwards produces a fallback that runs after a user's resolver crashed mid-handling.
  • The listing is async; the bootstrap is synchronous. The whole post-await commit therefore runs against a destination captured at request time (pmacs.window.commit_to), which preflights every precondition before invoking the callback — dired mutates handle state, prev, and paint long before it reaches anything that could refuse, so validating at display time is four mutations too late. Awaiting inside a commit is refused: a yield would restore the scope while the coroutine is still parked.
  • The scope swaps core.active_frontend, not just an override — pmacs.window.buffer()'s no-arg arm reads the ambient active buffer directly, so dired's prev capture would otherwise follow whatever frontend happened to be dispatching. The override also exists, and is load-bearing in exactly one case: a commit reached from inside an interactive command, where the origin would otherwise outrank the ambient value. Bite-testing found N4 green without it.
  • replace_active_buffer does not drop the startup scratch buffer, despite its doc comment having claimed so for as long as it has existed. Its body is one switch_active_buffer call. The comment is corrected here; changing the lifetime is separate work.
  • Stage 1b is the named remainder: compile binding + Cargo defaults, LSP spawn guidance, welcome buffer.

• Terminal configuration + copy mode arc — COMPLETE (docs/terminal-config-and-copy-mode-framing.md rev 4; Stage 1 #173, Stage 2 #178; no protocol change in either, still v20). Stage 1 ships profiles, scrollback, a per-terminal configurable escape key and the C-c t opener; Stage 2 ships copy mode — M-x terminal.copy-mode / C-c C-t.

  • The snapshot MATERIALIZES into an ordinary buffer. That is the arc's organizing decision: isearch, motion, selection and the kill ring work with no new substrate, and "keys must not reach the child" dissolves structurally, because the transport arm keys on is_terminal(buffer_id) and a snapshot is not a terminal. The dispatch-shadow count therefore stays at six.
  • prune reacts to buffer removal rather than causing it — it filters on !registry.contains(buffer_id), so a child exiting does not remove the terminal buffer. That is what makes on_removed a sound teardown hook, and why a finished command's output stays readable.
  • Ownership means "in our own handle table", never found-by-name (dired's F7 rule, re-learned here): snapshot writes use bypass_intercept, so adopting a same-named foreign buffer clobbers user data. Snapshot identity is keyed by comparing buffer handles in an array — BufferIdLua implements __eq but each wrapper is a distinct table key, so comparison works and hashing does not.
  • Profiles are a raw Lua table, joining pmacs.lsp.config and pmacs.pair.sets, because ConfigValue is four scalars with no table kind. The two open-time settings resolve through the global chain (they are read before the identity buffer exists); only terminal.escape-key resolves per buffer, and its cache lives on TerminalSession so its lifetime is the terminal's — value_epoch alone is not a sufficient key, because it does not advance when focus moves between terminals holding different buffer-local values.
  • Criterion 17 is deliberately unpinned, and its bite is now stated correctly. A real semantic frontend proving neither copy is mutated needs the actual GPU binary (the optimistic apply exists only in pmacs-gpu/src/main.rs; the headless SemanticClient has no optimistic path), i.e. the a37 footing §5 warns about. After set_generated_contents the eventual test must look for unauthorized mirror mutation plus daemon refusal — divergence, not the "mutates both sides silently" the criterion originally specified, which can no longer happen and would pass for the wrong reason. A fix can invalidate a test that was never written.
  • Test instruments worth reusing: cat -v is the echo probe, because the screen rejects C0 controls before they reach cells so a raw echoed Ctrl-X is invisible; and such probes must count occurrences rather than test presence, because a single-character probe collides with the child's own banner text.

• PTY terminate diagnostic LANDED — #176 (merge bf8878f, 2026-07-26, one review round; docs/process-signal-tolerance-framing.md rev 4, after three framing rounds). Diagnostic only — no disposition changed. Every call that failed before still fails, with no state transition and no reap-ledger arming; src/process.rs is the only source file touched. It exists because the terminate EPERM flake is real and nothing yet knows why.

  • Why three tolerance rules were all rejected: each concluded something about a process from something that was not about that process. Rev 1 reasoned from an errno alone (EPERM means the caller lacks permission, not that the id was recycled); rev 2 from try_wait, which observes the spawned leader while a PTY signal targets -tcgetpgrp(...) — entities that diverge exactly when job control has moved the terminal; rev 3 from group-directed ESRCH, which proves only that the selected foreground group vanished. This is the reusable shape, not a Unix trivium.
  • Two facts that killed the original argument. group = true is rejected for PTY mode at spawn, so the reap ledger never applies to the PTY path at all; and the ledger's own comment saying EPERM "cannot happen for our own children" drops the entry for bounded growth, not as a ruling that EPERM means dead. A comment stating a belief is not the same as code enforcing it.
  • What ships: a failing kill now reports five separate facts — target source, target kind/value, spawn-time group, errno, and the leader's real try_wait state. The test seam injects the kill result only, never the observation, so the real ChildHandle::try_wait runs against the real child.
  • It is not "strictly additive". try_wait reaps and caches, so an exited child may be reaped earlier than before. That is safe only because portable-pty 0.9.0 returns a std::process::Child on Unix and delegates try_wait to it, so poll_one still sees the cached status — pinned by an exactly-one-terminal-event test rather than assumed.
  • Still open, and still the most likely real fix site: signal_target's read-then-kill of tcgetpgrp. All tolerance rules remain parked pending the evidence this diagnostic produces, as does terminate idempotence for an already-reaped process (a different failure, so a different PR).

• Lean 4 arc (Arc 8) — stages 1, 2, 3a, 3b, 4a, 4b ALL LANDED (docs/lean4-mode-framing.md; #160, #161, #167, #170, #179, #181). pmacs edits Lean 4: arborium-lean highlighting, a lean4 major mode, ⟨⟩ ⦃⦄ ⟮⟯ pairs, and a lake serve language server with a Lake-aware outermost root, a lazy toolchain probe, a one-shot lean --server fallback, and waitForDiagnostics. No protocol change in any stage (still v20).

  • Two of the four stages contained no Lean at all, and that is the arc's organizing rule: no PR mixes a cross-cutting substrate change with Lean feature content. Stage 2 made LSP server affinity per-project-root (ensure_server had been reusing one server across roots — a correctness bug for every language, not just Lean). Stage 3a added notification/response subscription seams to handle_server_requests, the single shared LSP event drain, plus pmacs.fs.canonicalize.
  • Two consecutive re-scouts found that rule broken by the stage being scouted — Stage 3 in round 4, Stage 4 in round 5, each time by a risk column that contradicted its own prose. The rule is not self-enforcing. Re-check every remaining stage's risk column at scout time.
  • A configured LSP root must be a canonical absolute path. It reaches file_uri_for verbatim and that URI is the affinity key, so one package opened by two spellings spawns two servers. Stage 3a's pmacs.fs.canonicalize is the primitive; it returns nil rather than a lossy path for non-UTF-8 input.
  • LspManager::stop on an already-terminal client strands it in ShuttingDown forever — server_is_live then counts it live so nothing rebuilds against it, and forget refuses it for not being terminal. Stopping a dead server is what makes it un-replaceable. Stage 3b works around it by dispatching on state (forget when terminal, stop when live); merely skipping the call leaves next_restart_at armed. The real fix is unframed substrate work.
  • elan shims lie: lake --version and lean --version can both fail ("no default toolchain configured") on a machine where Lean otherwise works, so command -v lake is worthless as a capability check. Lean acceptance is fake-server; live smokes must be PATH- and success-gated.
  • Stage 3b took six review rounds, and the same defect appeared four times: "the fallback silently doesn't happen," as no re-attach, then re-attach cleared by an unrelated buffer, then satisfied by the very server being replaced, then repairing one buffer while the rest stayed stale. Each fix was locally right; none asked what a global config swap invalidates. The durable lesson is to heal at consumption — the point where a stale record is handed out — not at the moment of the swap.
  • Stage 4a (the typed-edit consumer chain) MERGED as #179 (branch lean4-stage4a-typed-edit-chain, framing rev 8; it is part of the main anchor above). It is substrate only: builtin/runtime/typed_edit.lua owns the single buffer.after-edit subscriber and the single one-shot read, pair.lua becomes its first registered consumer, and tests/auto_pair_acceptance.rs is unchanged by zero lines (criterion 46, verified at the diff). No protocol change, no Lean content. The three decisions that turned out load-bearing rather than stylistic: consumers are called even when the record is nil (three existing auto-pair tests assert the non-event through it, and 4b abandons stale pending state on it); each consumer gets its own copy of the record, because pairing reads rec.char and a declining consumer could otherwise forge it; and the fan-out iterates a snapshot, because a consumer that registers a lower-priority one shifts itself forward under ipairs and runs twice.
  • Round 8's durable lesson: run_all_must_succeed does NOT abort the fan-out. src/hook.rs:332 collects each callback's error and continues to the remaining subscribers, marking only the run failed — so an uncontained throw inside a hook subscriber does not stop lsp.lua from flushing didChange. Two framing revisions asserted the opposite to justify a pcall. The guard was right and the reason was wrong, and by the time review caught it the wrong reason had been copied into a module comment, an acceptance criterion, a test comment, and the ledger. Correct the source a rationale derives from, not only the sites that quote it.
  • Stage 4b (the Unicode input method) MERGED as #181 (framing rev 9): a vendored 1,855-entry table generated from leanprover/vscode-lean4@17d1d08 by scripts/regen-lean-abbrev, plus a consumer registered on the Stage 4a chain at priority 50, ahead of pairing. A consumer cannot both edit and let a later consumer act on the same keystroke: the chain hands each consumer a copy of the record made before any consumer ran, so an edit invalidates every copy still to be used. The expansion therefore runs on a SECOND buffer.after-edit subscriber after the chain — which is how a pair character that terminates an abbreviation still pairs (\alp( → α()). And deferring work past a fan-out means owning which fan-out it belongs to: these fan-outs NEST, so a consumer between the expander and pairing that calls pmacs.hook.run re-enters the deferred subscriber while the outer chain is still mid-list, and the count that recognises this has to come from a MINIMUM-PRIORITY consumer — the expander is optional (a claim can stop the chain first) and a subscriber beside the deferred one is too late (the nested fan-out finishes inside the outer chain's subscriber). Its other durable facts: the table must stay an ORDERED SEQUENCE (equal-length ties resolve by source declaration order, which a pairs-iterated map cannot express); a generator round-trip check must re-read the BYTES ON DISK, because comparing in-memory strings cannot see an encoding applied by the write itself; and an expansion that SHRINKS the buffer must place the point explicitly, or every later self-insert is silently rejected and the editor looks dead.
  • Round 9 corrected three approved acceptance criteria by simulating the state machine over all 1,855 entries rather than re-reading the prose. Four review rounds over the text had not found them, because each named an example that reads as obviously right and is wrong only against the data.
  • Remaining: stages 5 (goal panel), 6 (#eval output channel), and 7 (module hierarchy) are framed but not scouted against current main.

• Inline math LANDED — #158 (docs/inline-math-slice-framing.md rev 3; merge 5aa9044). pmacs renders $…$ as typeset mathematics in the GPU frontend. No protocol change (still v20); the whole slice lives in pmacs-gpu, because pmacs-gpu depends only on pmacs-protocol and never on pmacs — a core-crate parser would have been unreachable from where rendering happens.

  • math_parse.rs → math_layout.rs → a ChunkSource::MathBox spacer chunk → per-glyph mini-buffers drawn at the shaped line's real baseline, with fraction rules as quads. Font is bundled Latin Modern Math (~717 KiB) under the GUST Font License — not OFL.
  • The v0 subset is narrow and deliberately so: Greek (34 entries), sub/superscript, and \frac. Everything else — including relations like \geq, fences, big operators, and all display math ($$…$$, \[…\]) — is a named deferral, and an unsupported command degrades the whole span back to source rather than rendering partially. In a real paper most inline spans still show source; that is the designed behaviour, not a defect.
  • Math is suppressed while the caret is inside its span, so editing always sees source. That gate reads the effective caret plus selection endpoints and is fed by three separate refresh triggers; it is the most delicate part of the slice.
  • Selection and search washes cover the whole box rectangle, not sub-ranges (sub-range washes are deferred).
  • TUI shows the LaTeX source unchanged. That divergence is recorded against COHERENCE.md §16, which audits the "no privileged frontend" rule.

• find-file LANDED — #162 (docs/dired-framing.md §10, Q#DR11; merge 2af1ab3; one review round). C-x C-f is the dired arc's Stage 0: pmacs previously had no discoverable way to open a file by path — no such command existed and pmacs.buffer.find_or_open had no interactive caller. Pure Lua in builtin/commands/default.lua, one keymap line, an 8-test dispatch-driven acceptance suite; no Rust, no protocol change. Two substrate facts it documents, both worth knowing before touching any minibuffer prompt:

  • Completion over files is flat and cannot be made hierarchical from Lua. A custom source function is called with zero arguments (minibuffer.rs:591) and runs synchronously outside any coroutine, where Handle:await() raises — so it can neither see the input to re-root on nor list a directory. Only the Rust CompletionSource::Files { root } can list, and it is single-directory and 1024-capped.
  • A selected candidate SHADOWS typed text. recompute_candidates sets selected = Some(0) whenever the list is non-empty (minibuffer.rs:372-377) and resolve_accepted_value returns the candidate over the typed contents (:564-574). So free-text accept fires only when the input filters every candidate away — for basename candidates under a subsequence filter, when it contains a /. This applies to M-x and switch-buffer too. Consequences are pinned as decisions, including the hole where a new bare name that is a subsequence of an existing entry opens the existing file, and the empty-input case (fuzzy_score gives Some(0) for an empty needle and ties break lexicographically, so dotfiles lead).
  • Also: get_or_load_buffer computes a normalized path but loads from the raw one (editor_core.rs:842-856), so a ~/… path dedups against an open buffer yet fails to load one that is not open — find-file expands the tilde Lua-side. Loading through the normalized path is a named deferral.

• dired Stage 1 — the directory view — LANDED — #165 (docs/dired-framing.md §0, S1-1…S1-12; merge c8ec8f3; one review round). pmacs now has a directory surface: C-x d / C-x C-j open a read-only listing, one buffer per directory named *dired:<canonical path>*, with a dired major mode whose mode-scoped keymap carries RET/f, ^, n/p, g, q, s. Protocol unchanged at v20. Stage 2 (marks and operations) and Stage 3 (wdired) each still need their own framing; the frozen fixture shrinks after Stage 3.

  • The Rust is confined to two things: a per-entry-tolerant read_dir (ReadDirTolerance {Fatal, PerEntry} → FsDirListing {entries, errors}), because read_dir_blocking fails a whole listing on any of five per-entry conditions and the tolerant wrapper its own module doc delegates to package authors cannot be written in Lua (one error value, no partial vec); and editor_core::normalize_buffer_path becoming pub, exposed as pmacs.path.canonicalize. Only non-UTF-8 names stay fatal — byte-preserving paths would be needed. The Lua result shape keys on errors.is_some(), so the bare array the frozen M8.2 fixture consumes with ipairs is untouched.
  • Exposing a core normalizer beat mirroring it in Lua. A Lua mirror would have been a second canonical form — the same class of bug as the five tab-width constants (#137). Applies to any future Lua-side path reckoning.
  • A fixed-width column must be fixed-width for every input. The exported pmacs.dired._layout (MARK 0, KIND 2, PERMS 3–12, SIZE 13, MTIME 24, NAME 41) is the contract Stage 3 reads offsets from, and %10d overflows at ≥10 GB, silently shifting every column right of it. Sizes now fall back to a width-clamped magnitude (K/M/G/T/P/E).
  • An ambient action must be gated on the buffer it assumes. A revert's cursor re-seat settles a tick or more later, by which time the user may have switched buffers; the paint names its buffer and is safe, but seating is ambient. This is the buffer-level instance of the rule below that interactive origin does not survive an await.
  • A failure IS an answer — don't probe first. Kinds are lstat-based in both read_dir and stat, so nothing in an entry says whether a symlink points at a directory. RET tries to list it and treats the failure as the answer; an explicit probe was a second full read_dir, so a descent listed twice.
  • Unbounded per-entry error collection needs a cap when nothing cancels the work. A dired listing carries no supersede key, so cancellation was never the backstop the tolerant loop implicitly relied on (READDIR_MAX_CONSECUTIVE_ENTRY_ERRORS = 1024).
  • This is the first builtin with mode-scoped keys (#129's first non-detection consumer), which broke the pre-existing describe_key_identifies_every_default_binding: it asserted every binding resolves through describe.key context-free, which held only while the modes table was empty. It now sets the effective context per binding and explicitly clears the mode for global ones, because a leaked mode legitimately shadows a global chord of the same name (dired's RET shadows edit.newline-and-indent), plus a floor assertion that at least one mode-scoped binding exists.
  • A dedicated panel does not carry its dedication across a descent — the framing expected it to. display_buffer never replaces the buffer in a slot dedicated to another one; it discards every side-specific parameter and falls back to the document window (Q#BP3 2.iii), and the exact-window arm errors. Dired does not unpin the user's panel; both arms are pinned.
  • Smaller facts worth knowing before touching this code: a path-backed buffer's name is its full path, not its basename, which matters for any name assertion; pmacs.buffer.kill (not remove) redirects windows off a doomed buffer first, so dired.kill-when-opening kills after the replacement is displayed; ownership is checked against the handle table only, never the buffer name; and C-x d takes no completion source on purpose (with one, RET on an empty field opens whatever sorts first, and RET-where-you-are is the gesture the binding exists for — the field is prefilled instead).
  • Verification at merge: 1,832 default + 2,009 CRDT library tests; dired acceptance 25 + 25 CRDT; the frozen m8_1 10 / m8_2 15 / m8_3 32 unchanged, which is the additivity gate for the read_dir change; M4 121; required GPU 155; isolated-XDG_CONFIG_HOME workspace sweep 3,205 across 93 suites. 15 claims bite-verified.

• Canonical main is protocol v20 (SUPPORTED=[6..=20]; v16 = ThemeFacts, v17 = FontFacts, v18 = StatuslineSegments, v19 = terminal frames/events, v20 = the GPU initial-target semantic bootstrap family). Bottom-panel Stage 2B-1's in-review schema is v21 (SUPPORTED=[6..=21]), but its production daemon deliberately advertises v20: the handshake is server-first, so advertising 21 would make shipped v20 GPU/TUI clients reject before AttachRequest. Stage 2B-3 owns compatible production activation.

• Bottom panel Stage 1 (window placement + TUI side windows) LANDED — #155 (docs/bottom-panel-framing.md rev 4; merge e745068; two review rounds). No protocol change (still v20). Arc 7's substrate: pmacs now has Emacs's display-buffer + window parameters, and a buffer can be displayed in a fixed-height window pinned to the bottom of the frame that feature code targets by policy instead of by stealing the selected window.

  • src/window.rs: WindowParams { side, fixed_rows, dedicated } plus implementation-owned quit_action / origin_document (Lua reads them, set_params refuses them); MIN_WINDOW_OUTER_ROWS = 2; Layout::compute(area, fixed) subtracts fixed children before dividing the remainder by weight, preserving last-flexible-takes-the-remainder, so a tree with no fixed leaves computes byte-identically to before.
  • Layout::compute has TWO production callers, and both must feed the same shared panel_fixed_rows map: window_placements and src/overlay_paint.rs's peer-presence pass, which derives its own text-area Rect and never routes through the first. Leaving the second on unfixed geometry paints every peer cursor at the row it would occupy with no panel open.
  • The minimum is recursive (subtree_min_rows: horizontal splits sum, vertical splits max). "Two rows at the root" does not give each nested leaf two rows. interactive_min_rows is the same recursion over the user's window.min-height preference, and applies to drag/keyboard resize ONLY — the layout pass and frame-resize reconciliation use the structural floor, so changing a preference can never invalidate an existing layout.
  • Hiding a panel is a durable state transition, not a per-frame effect (EditorState::reconcile_panel_layout): it moves focus out and releases the terminal controller, because the terminal resize path merely returns on zero content without releasing. It runs after attach/resize/display/ close and defensively before input dispatch, terminal sync, and paint.
  • FrontendView gains panel_capable, frame_geometry (None = unknown, never the GPU attach request's permanent 24×80 placeholder), and derived panel_hidden — each spelled explicitly at every construction site, preserving folding's non-Default discipline.
  • EditorCore::primary_document_window is the Q#BP14 projection seam; display_buffer is Phase 1 of the display transaction (exact target → side affinity → ordinary reuse, with option-valued height/dedication); the Lua layer owns Phase 2 (activate → hook → reconcile → revalidate → final-focus matrix).
  • Optimistic input is gated per WINDOW, not per buffer: dispatch_idle_for returns false whenever the acting frontend's active window is a side window. Marking the panel's BUFFER round-trip would be wrong — round_trip_buffers is global by BufferId, so it would disable optimistic apply for another frontend editing that buffer as its document.
  • Jump entries are per frontend and carry their origin WindowId; a stale side origin is SKIPPED, because degrading it to an active-window switch is exactly the duplicate-panel corruption the arc removes.
  • pmacs.window.display / display_file / quit / panel / params / set_params / resize / display_target, plus builtin/runtime/window.lua (window.panel-height, window.min-height, C-x ^ / C-x C-^). Adopters take display = "current" | "panel"; Stage 1 default is "current" and Stage 3 flips it.
  • The divider is the upper subtree's existing mode-line row — no row added or consumed, ui.divider restyles every exposed segment of one boundary, and drag state is HashMap<FrontendId, _> so frontends cannot steal each other's gestures.
  • open_initial_target now shares one resolve_target_buffer + exact-window install seam with display_file, and reasserts into a document window after hooks (a startup hook can now create a panel).
  • Final gates: 1,817 default + 1,994 CRDT library tests; the new bottom_panel_stage1_acceptance 46; kill ring 30; compile 67; M4 121; required GPU 152; initial-target 14 CRDT; all three vterm suites; folding Stage 2 48. All 12 CI checks green at merge.
  • Stage 2 (the GPU panel band) is FRAMED, and its first two slices have LANDED — docs/bottom-panel-stage2-framing.md rev 6, four framing review rounds, no open framing items; the rev-5 implementation split was explicitly approved 2026-07-27 and rev 6 records PR #184's server-first compatibility and gate correction. It reserves protocol v21 and ships as four serial implementation slices: 2A classified census routing + per-window painter extraction (no wire change, #177), 2B-1 the wire (#184), 2B-2 the daemon projection and epoch machine, then 2B-3 the GPU band, compatible v21 activation, and the negotiated panel_capable flip. Production attachment remains v20 through 2B-2. Parent acceptance 37–55 remains authoritative. Stage 3 is the adopter default flip. 2B-2 is the next slice, and it branches from 0442d78 or newer.
  • The §1.3 census is CLASSIFIED, not uniformly redirected. Only the Projection class (#1–#12, #21–#22) routes through primary_document_window; focus/input (#13–#15, #23), focus chrome and surface-routed (#16–#19), and focus/session (#20) keep their own authorities. Rerouting them breaks remote-op validation and application, DispatchIdle, presence, focused search/menu/completion routing, and terminal bell ownership.

• Bottom panel Stage 2B-1 (the reserved v21 wire) LANDED — #184 (merge 6bee09d, 2026-07-28, two review rounds plus a gate-found follow-up; all 12 checks green on reviewed head 5539b6e). It adds no producer, no consumer, and no capability: panel_capable is still false for every semantic session, so nothing about it is user-visible. What it establishes is durable:

  • Schema support and production advertisement are separate facts. SUPPORTED is now 6..=21; the daemon's unsolicited Hello still says 20. This is not a hedge — the handshake is server-first, so a shipped v20 frontend rejects a Hello { protocol_version: 21 } before it can send an AttachRequest. Bumping the advertised version is therefore an incompatible act on its own, independent of whether any new message is ever sent. A real-daemon acceptance emulates that exact rejection point and then requires the attachment to reach its initial grid. 2B-3 must ship a compatibility-preserving activation mechanism; it may not simply change the unsolicited Hello to 21.
  • One shared grid validator, split along a stated boundary (pmacs-protocol/src/wire_grid.rs). Shared: checked area, the visible-cell bound (262,144), cell count, cursor bounds, glyph legality, wide-continuation topology, the 8 MiB aggregate glyph-byte budget, and the attachment rejection. Terminal-only: the 512 per-axis PTY caps, title/process metadata, selection spans, and the at_bottom == (scroll_offset == 0) coupling. Per-axis caps are a WireGridLimits parameter, not a constant, because a panel does not inherit them — a 4K surface at a small font is legitimately wider than 512 columns, and the area bound is what keeps the encoding inside the transport budget. The attachment rejection is deliberately shared: panels render no attachments either, so classifying it terminal-only would let a panel ship a cell no frontend can paint.
  • PanelFramePayload::Absent is authoritative, and silence is not. The receiver retains its last valid frame, so a close or a hide must send Absent explicitly or a stale band stays on screen indefinitely. Validation is likewise atomic — a bad frame is rejected whole and the previous valid frame is retained.
  • Two epochs, answering two different questions. panel_epoch is opaque and monotonic per frontend: stable across ordinary frames of one continuously present window/buffer, and moved on buffer replacement, new side-window creation, and every Absent → Present transition — which is what stops a stale PanelPointer from addressing a reopened panel as if it were the old one (Q#BP16). geometry_epoch answers a frontend declaration and moves whenever the frontend declares new effective cell geometry, including a font or scale change that leaves CellSize identical — exactly the case daemon-side value dedup cannot see (Q#BP2S1).
  • PanelFrame carries an explicit buffer_id (review round 1) and its focused bit is presentation and focus-chrome routing only (Q#BP14b) — the keys decision remains DispatchIdle (Q#BP14a).
  • The frontend half is FrontendEvent::{FrontendCellGeometry, PanelResizeRows, PanelPointer}, gated in both directions, with each extended enum byte-pinned on its own previous final variant so the v6–v20 encodings are provably unchanged.
  • A version bump is not done until every ratchet that pins the old version moves. The full gate — not review — found that the statusline and Vterm Stage 3 ladders still pinned v20 and rejected v21, in both structural and real-headless-probe form. Grep for the outgoing version across tests/ before calling a bump complete.

• GPU initial target LANDED — #148 (docs/gpu-initial-target-framing.md rev 3; merge 0dd16a5; two review rounds). pmacs --gpu [--socket NAME|PATH] FILE transports exact Unix path bytes plus launcher cwd to the managed GPU client. Protocol v20 adds a semantic-session SessionBootstrapRequest after AttachRequest and an appended InitialTargetResult readiness barrier; v6–v19 wire encodings stay pinned. The daemon resolves the path lexically, deduplicates or loads/creates it in the authenticated frontend's view, runs the established load/switch hooks, upgrades the buffer for CRDT, and publishes every target-side CRDT upgrade to existing grid replicas before readiness. Semantic replicas receive a publication only when displaying that buffer, so a second target launch cannot switch an existing GPU window; one dead peer cannot fail the new session. Failed bootstrap writes a bounded result, shuts down the socket, removes provisional state, and restores the ambient active frontend. Any stale event from an uninstalled session is dropped before state access. Existing no-target managed launch, direct attach, TUI, and legacy protocol behavior remain intact. Folding Stage 2 integration: fold projection at attach is selected from the same negotiated semantic_render bit the target bootstrap uses (grid collapses, semantic/GPU stays source-line pending Folding Stage 3). Final gates: 1,815 default + 1,992 CRDT library tests; target + invocation gates 14/14 CRDT each; Folding Stage 2 48 CRDT; M4 121; required GPU 152; Vterm Stage 3 5 default + 7 CRDT; isolated-config workspace sweep 3,334 across 88 suites; two concurrent real Wayland/Vulkan GPU windows stayed on distinct target buffers after the second attach. All 12 CI checks passed.

• Folding Stage 1 (headless fold engine) LANDED — #142 (docs/folding-framing.md rev 5; merge c49a8c7; three review rounds, round 3 clean). Arc 6's engine — instance-side and headless; no frontend renders a collapse yet (that is Stage 2). No protocol bump.

  • src/fold.rs: a per-buffer FoldStore of byte ranges attached as a translating/dropping View (the BufferStyleSpanTranslator pattern — it translates strictly-inside edits and DROPS boundary-crossers, provenance-blind); FoldRegistry/SharedFoldRegistry = Rc<RefCell<HashMap<BufferId, {Arc<Mutex<FoldStore>>, ViewId}>>> (the SyntaxRegistry per-buffer model), held on both EditorCore (src/editor_core.rs:223) and EditorState (src/editor.rs:110). Containment is start-exclusive, end-inclusive (start, end]; the stored range is [end of head line, end of last hidden line] (NOT the ByteRange struct doc's [start,end)).
  • Structural source: nearest enclosing block-like node ≥2 source lines → resolve introducer↔body → derived head line (the line immediately above the first hidden line, so wrapped signatures / where clauses stay visible) → closer-aware tail (a closing-delimiter line stays visible, e.g. } else {). Stale/absent parse tree refuses.
  • The six EditorCore edit primitives call unfold_before_point_edit first (command-path pre-edit unfold, keyed on the active frontend's point). Interactive Lua-command unfold (yank/query-replace/comment) is a Stage 2 obligation; CRDT-origin unfold is Stage 3.
  • src/lua_bindings/fold.rs (install_fold): pmacs.fold.* data API (explicit buffer, no ambient resolution, matching #127) + interactive commands on the Emacs hideshow C-c @ prefix set; builtin/runtime/fold.lua.
  • src/semantic_render.rs: fold_state_msg PRODUCES FoldState (semantic/GPU sessions) — authoritative-empty, diff-suppressed, per-session baseline reset on BufferSnapshot (the #120 stale-mirror trap class; the GPU fold-mirror clear-on-snapshot is a named Stage 3 obligation).
  • Durable lesson (round 2): after wiring a cleanup into a production hook, PIN IT THROUGH THE REAL PATH — a direct-call unit test misses the wiring (falsify by revert).
  • Stage 2 (grid/daemon collapse) LANDED — #149 (merge 6ed4fe9; five review rounds; docs/folding-stage2-framing.md rev 4). Its load-bearing reframe: the TUI had no non-identity source-line↔display-row map (view_top + row was baked into ~13 sites), so Stage 2's spine is src/fold_view.rs's VisibleLineMap — derived from the fold store plus a window's line offsets, never stored — that the render loop, gutter, diagnostics, caret, selection, peer presence, viewport/scroll/motion, and the mode-line indicator all route through, plus the interactive-Lua unfold widening. Threaded as Option<&'a VisibleLineMap> on a lifetime-bearing Viewport<'a> that stays Copy. Design points the review rounds forced, each a trap for Stage 3:
    • the map's unit is a merged hidden component (overlapping or adjacent intervals unioned, keeping the earliest visible head), not a fold — folds may cross, and an inner/later fold's own head can be hidden;
    • instances are per rendered window and per command/event operation, never per frame; a command's map follows the operation's TARGET window (a wheel event names a pane without activating it);
    • fold projection is per-frontend (FrontendView.fold_projection, set at attach from the negotiated semantic_render bit) — shared EditorCore motion would otherwise make a simultaneous unfolded GPU session's cursor skip lines it still displays;
    • a hidden cursor normalizes by position, not row, and set_view_top clamps in the setter rather than being repaired at render time;
    • the interactive-Lua unfold keys on the post-intercept edit site — a managed buffer intercept may legally relocate the op. No protocol bump. Stage 3 (GPU) is next and has no framing yet; its named obligations are GPU collapse at TUI parity, caret/hit-test fold-awareness, the BufferSnapshot fold-mirror clear (parent R2-4 — the #120 trap class), CRDT-origin / GPU-optimistic interactive unfold (parent R2-3), and flipping FrontendView.fold_projection to true for semantic frontends.

• One-command GPU invocation LANDED — #141 (docs/gpu-invocation-framing.md rev 6; merge 63fbc66; two implementation reviews). The additive public path is pmacs --gpu [--socket NAME|PATH]; bare pmacs [FILE] remains the TUI. Root owns the CRDT gate, socket resolution, sibling-regular-file GPU discovery/PATH fallback, and GPU outcome. The separate pmacs-gpu binary owns connect-or-start, a five-second / 50-ms retry window, pre-winit event buffering, daemon process-group/stdin/stdout/stderr isolation, and named child reaping with explicit ownership handoff. Direct pmacs-gpu --attach RAW_PATH remains strict, is documented as advanced, and never auto-starts. No protocol change. The initial macOS/LuaJIT CI run exceeded an unrelated outline performance threshold (147 ms / 100 ms); the complete failed-job rerun passed all twelve checks before merge.

• Config registry LANDED — #127 (docs/config-registry-framing.md rev 3; merge 2e37c04; two review rounds). pmacs.config is the typed, introspectable options registry the backlog ranked first, and it closes the "config-registry-blocked" deferrals below. It was built as a PARALLEL LANE alongside vterm in a sibling worktree; the files were assigned per-lane up front and the rebase had zero conflicts.

  • Third registry beside CommandRegistry/HookRegistry (src/config_registry.rs), same R42/R50/duplicate-rejection/ SourceLocation vocabulary; Lua surface in src/lua_bindings/config.rs. No protocol change (still v18) and ZERO changes to src/editor.rs.
  • An override is ALWAYS stored, even when equal to the value it shadows; only value_epoch and listener dispatch key on effective change. The "equal-value set is a no-op" reading silently voids a buffer-local pin: nothing is stored, and a later global set flips the very buffer the user pinned.
  • Two scopes: global and buffer-local. get(name, buf) resolves local → global → default; get(name) resolves the GLOBAL CHAIN ONLY and never consults an ambient buffer. Per-language and per-project are patterns (a hook calling set_local), not scopes the registry knows about. Mode keymaps now resolve through #129, but pmacs.config deliberately remains global + buffer-local.
  • Buffer-locals live in a registry side table purged at after_buffer_removed, beside the keymap purge.
  • Listeners: commit → snapshot → drop the borrow → re-enter Lua; a raising listener is logged without blocking the rest or rolling back; a depth bound turns a cycle into a pointed error. Explicit dispose only — there is no MetaMethod::Gc anywhere in the codebase, and GC timing differs between the two Lua backends.
  • StartupOnly freezes off the existing InitCompleteFlag at write time (which is why no editor.rs call was needed). In --lib builds set_init_complete never runs, so a post-freeze test must flip the flag explicitly or it passes vacuously.
  • Adopters own their own define, so SourceLocation names the owning module: editing.auto-pair (pair.lua, read against the typed edit's SOURCE buffer), editing.trim-on-save (editops.lua, read against the buffer being saved), autosave.interval-ms (autosave.lua, re-read per tick). The migration wrappers keep their legacy coercion — the registry is strict, the legacy setters stay lenient (trim_on_save("yes"), interval_ms(1500.7)).
  • M-x describe-setting renders into *help*.

• Mode system wiring LANDED — #129 (docs/mode-system-wiring-framing.md; merge b4b925d; one review round). The existing mode-keymap substrate is now live without a protocol change.

  • Buffer.major_mode: Option<String> owns the single major mode. The detected language name initializes it once on buffer.after-load, before grammar gating, so server-only languages work; switches never rewrite it. Explicit overrides and clears survive switches. A future reload that fires after-load re-detects, and explicit mode state is not session-persisted.
  • Dispatch borrows the zero-or-one mode through Option<&str>::as_slice() and &[&str]: no hot-path mode allocation. Resolution remains buffer-local → mode → global, and registry/keymap borrows end before Lua command invocation.
  • Lua surfaces: pmacs.buffer.major_mode / set_major_mode and pmacs.editor.active_modes. pmacs.describe.key, pmacs.help.show_key, and followed percent-encoded @mode: links use the same effective context; pmacs.keymap.lookup remains raw-global.
  • The built-in mode statusline provider reads ctx.buffer, so passive splits render their own mode. Real-daemon acceptance covers all ten framing criteria across both Lua backends and Linux/macOS CI.

• Modeline language detection LANDED — #132 (docs/modeline-detection-framing.md rev 2; merge 1dd47fc). Fresh loads scan bounded Emacs -*- mode: ... -*- and Vim/Vi ft= / filetype= modelines without evaluating file content, normalize common editor aliases, and give explicit modelines precedence over inferred language.

  • builtin/runtime/syntax.lua owns one per-buffer fresh-load decision: modeline → bundled grammar extension → LSP filetype extension → exact filename → shebang. Syntax, initial major mode, pairing, comments, and LSP all reuse that pin; LSP retains its independent backing-path guard.
  • Editing a modeline or shebang does not switch an attached parser or make language-aware consumers diverge. Close/reopen re-evaluates changed file metadata. Explicit post-load major-mode overrides remain independent.
  • Bounded valid unknown names remain passive major modes without starting an unavailable parser or server. All thirteen framing criteria are covered on LuaJIT and Lua 5.4. No Rust or protocol surface changed; protocol stays v18.

• Syntax-highlight / language-detection side-quest (#114–#118) LANDED — a one-shot arc built in sibling worktrees off main while the user's themes lane (theme-faces) ran concurrently in the shared checkout. All merged. What shipped:

  • Grammars (crate::syntax::BUILTIN_LANGUAGES): every LSP-configured language now has one — cuda, bash, dockerfile (via the ABI-current tree-sitter-containerfile, NOT the dead tree-sitter-dockerfile which pins tree-sitter ^0.20), make, cmake, python, go, javascript (+jsx), typescript (+tsx), toml, zig.
  • Detection chain and pin (builtin/runtime/syntax.lua): modeline → grammar extension → LSP filetype map → filename → shebang. User-extensible Lua surfaces include pmacs.parse.modeline_aliases, .shebangs, .filenames, language_from_modeline, language_from_shebang, language_from_filename, and the pinned buffer_language. builtin/runtime/lsp.lua delegates to that shared decision after enforcing its backing-path requirement. Grammar name MUST equal the pmacs.lsp.config.<name> key. A buffer keeps its pinned language across edits/switches; close/reopen performs a fresh bounded inference.
  • LSP configs added: dockerfile (docker-langserver --stdio), cmake (cmake-language-server, config via init_options.buildDirectory="build" — it does NOT pull workspace/configuration). Make has no server.
  • Substrate: LanguageEntry.highlights_query and .locals_query are &[&'static str] fragments joined base-first (cuda over c/cpp; ts over js/jsx). Since locals-query processing #134, settle compiles the grammar's LOCALS_QUERY, resolves Tree-sitter's scope/definition/value/ reference conventions into sorted LocalFacts, and stores them beside each layer's tree/query. Work runs once per fresh bundle and only when the highlight query asks about local; viewport rendering remains bounded. Both TUI and semantic/GPU producers evaluate #is?/#is-not? local through the shared capture walk. Non-shadowed JS/TS builtins are restored; shadowed definitions/references keep ordinary variable styling.

• Multi-language injections (#122) LANDED — the direct continuation of the #114–#118 highlight arc; four review rounds, framing docs/multi-language-injections-framing.md (Q#IJ1–IJ11). A buffer can now hold more than one language: ParseTreeBundle holds Vec<Layer> (root + injected children, depth-ascending, installed atomically so the existing StyleGate + highlight-cache Arc gates still work). The parse worker builds child trees off the static BUILTIN_LANGUAGES table (lazy load preserved) via set_included_ranges — child node offsets are ABSOLUTE, so injected spans are buffer-coordinate-native; settle resolves each layer's highlight query (SyntaxRegistry::resolve_layer_queries). First consumers: markdown fenced code + markdown_inline (retires the M9.7 block-only floor). New substrate: LanguageEntry.injections_query; default_injection_aliases + SyntaxRegistry::injection_alias_snapshot (case-folded fence names, Lua-extensible via pmacs.parse.injection_aliases, snapshotted into ParseRequest so the worker never touches the Rc registry or Lua); ParseTreeBundle.injection_capped (the 4096-layer backstop, surfaced once/buffer via pmacs.error at settle); compute_highlight_spans_for(query, tree, source, local_facts, range) (per-layer); the wire flatten_layer_spans event-sweep → DISJOINT effective spans (deeper / later-sibling / narrower wins, keyed by (layer_index, capture_order)); GPU spans_from_segments + source_color_at fold. Two findings to keep: injection ranges exclude only NAMED children (anonymous tokens ARE the injected text — excluding them shreds a block inline node; matches tree-sitter-md's own splitter), and the wire flattener runs over the WHOLE buffer via the file-style summary, so it must be an event sweep, not O(spans²).

• JSON + YAML grammars and language servers (#123) LANDED — bundled ABI-current tree-sitter-json / tree-sitter-yaml cover .json, .yaml, and .yml; the existing injection engine now highlights YAML frontmatter and JSON/YAML fences. Default external LSP configs are the pinned vscode-json-language-server provider and yaml-language-server; configured settings are pushed after initialized, which also supports push-model servers. The fake-server delivery proof and PATH-gated live JSON/YAML provider smokes cover the configuration contract. .jsonc / .json5 remain a deliberate follow-up because the JSON grammar is strict.

• Compile-mode (Arc 5 stage 1, #113) LANDED (2026-07-14, 7 rounds; framing docs/compile-mode-framing.md rev 13). compile.run streams /bin/sh -c "exec 2>&1; <cmd>" into an intercept-read-only *compilation* buffer via a Lua ANSI parser; once-per-newline error rules; unified error.next/error.previous (M-g n/p, C-x `, M-!). Substrate other code can use: ProcessSpec.stdin/group (group lifecycle: reap ledger, in-drain enforcement, cancellable poll readers), buf:revision(), jump_back fires buffer.after-switch, pmacs.errors.claim, AnsiParser::finish() (observable reset), and the style-overlay stack: buffer-attached BufferStyleSpanTranslator (once-per-edit, fragment-preserving), render-only window overlays with identity-deduped Window::ensure_overlay + clone_for_split, idempotent atomic handle:dispose().

• Editing-conveniences pack (editops, #111) landed (framing docs/editing-conveniences-framing.md rev 6). goto-line, case ops, transpose, zap-to-char, line move/duplicate/join, region sort/reverse/dedupe, delete-trailing-whitespace + opt-in pmacs.editops.trim_on_save. Substrate: pmacs.killring.kill_range / break_chain([fid]) / arm_kill_prompt+commit_kill_prompt (marker lifecycle), and the origin-guard pattern for chain-sensitive minibuffer commands.

• Auto-pairing (#110) landed — Arc 2 COMPLETE (framing docs/auto-pairing-framing.md rev 6). BUILTIN_PAIR_CHARS in pmacs-protocol leave both frontends' optimistic classifiers; builtin/runtime/pair.lua loads BEFORE lsp.lua (first-didChange ordering contract); one-shot typed-edit provenance via pmacs.editor.take_typed_edit() (buffer-revision postcondition, Q#AP9). Substrate: buf:path(), pmacs.lsp.buffer_language(buf), PMACS_FAKE_LSP_CHANGE_SINK, TestDaemon::spawn_with_config.

• Themes (Arc 4) stages 1–3 LANDED; Arc 4 COMPLETE ON main.

  • Stage 1 (#120, docs/theme-faces-framing.md rev 9): named UI faces as reserved ui/ui.* theme entries; transactional split syntax/face epochs; protocol-v16 ThemeFacts; snapshot/baseline symmetry; store-sourced diagnostic-count freeze.
  • Stage 2 (#124, docs/gpu-set-font-framing.md rev 5): pmacs.gpu.set_font and authoritative protocol-v17 FontFacts; frontend-local family resolution, live font reload/reflow, and visual-run caret geometry.
  • Stage 3 (#125, statusline-segments, docs/statusline-segments-framing.md rev 3): composable strict pmacs.statusline providers; borrow-released per-window evaluation with failure latches; legacy-preserving TUI composition; a pure built-in LSP provider; dynamic modeline faces; protocol-v18 StatuslineSegments; authoritative-empty/snapshot symmetry; and atomic GPU validation, face resolution, shaping, clipping, and cache invalidation. Acceptance 1-27 is implemented. Final gates: Clippy clean; 1,619 default + 1,793 CRDT library tests; 7 default + 8 CRDT feature acceptance; 114 M4; 109 required GPU; one-invocation workspace sweep 2,718 passed across 78 suites (19 ignored, basedpyright filtered); git diff --check clean. Stage 3 landed as #125 and completed Arc 4 on main.

• Vterm Stage 1 terminal core LANDED ON main — #126 (docs/vterm-framing.md rev 5; merge 643d1e1).

  • Implementation commits: bbc1f33 (Stage 1), 962944b (Darwin signal normalization), first-review fixes f0a235f, 28f2e6c, bf972a7, and second-review hardening 9797ada; reviewed feature head fc4e0ce merged through PR #126, https://github.com/levineuwirth/pmacs/pull/126.

  • AnsiParserProfile::{LineOriented, FullScreen} preserves compile/REPL behavior while terminal PTYs emit the full cursor/mode/device operation set. src/terminal/{screen,input,session}.rs owns the state machine, encoders, and lifecycle registry.

  • Public session seam: owned strict TerminalSpec; owned TerminalSnapshot; TerminalProcessState; and SharedTerminalManager = Rc<RefCell<TerminalManager>> with open/is_terminal/process_id/snapshot/tick/send/resize/terminate/prune/ shutdown. Stage 1 snapshots are context-free; Stage 2 adds per-view state without a second screen.

  • EditorState tick order is supervisor → terminal-owned PID drain/prune → process.after-tick. Terminal IDs are not exposed through pmacs.process; ordinary Lua/LSP/MCP ownership is unchanged. Terminal identity buffers are pathless, clean, empty, round-trip, and guarded read-only at every rope/CRDT/history mutation boundary.

  • Acceptance 1–14 is mapped in the framing. The real PTY bite splits ESC/CSI writes, observes alternate-screen cursor addressing, blocks and resumes through raw send, restores the main screen, and pins final output before exact PID/outcome annotation. One-row annotation visibility, TERM-ignoring shutdown, spawn rollback, buffer-kill prune, and immutable empty CRDT bootstrap are pinned.

  • Review round 1 added typed IND/NEL/RI with margin-correct screen behavior, defaults absent TERM to xterm-256color, makes shutdown liveness acceptance portable with kill(pid, 0), and preserves custom tab stops on resize. Review round 2 rejects C0/C1 controls before they enter screen cells, preserves the released button code in SGR mouse reports, removes dead screen paths, and clears stale round-trip state during prune. Stage 2 now uniquifies default terminal buffer names transactionally.

  • Exact CUU/CUD and out-of-range DECSTBM clamping, combining across controls, xterm alternate-screen details, legacy non-SGR mouse, printable ASCII and CSI-dispatch allocation fast paths, and scrollback-cap naming are explicit post-arc deferrals in the framing.

  • Final from-start rerun after review round 2: Clippy clean; 1,661 default + 1,837 CRDT library tests (3 ignored each); 9 default + 10 CRDT vterm acceptance; M4 114 passed (3 ignored, 1 filtered); required GPU 109; workspace 2,769 passed across 79 suites (19 ignored, 1 filtered); diff check clean. scripts/bite HEAD^ src/terminal/screen.rs --test vterm_stage1_acceptance terminal_cells_reject_child_control_characters is a clean behavioral bite. The parser dispatch has its independent clean behavioral bite; the original main/crate-root bite remains explicitly weaker compile-time API evidence.

  • Stage 3 GPU/protocol LANDED ON main — #135 (merge cac4961; docs/vterm-framing.md Revision 9, criteria 28–37). Protocol v19: InstanceMessage::TerminalFrame (discriminant 26, daemon-gated) plus FrontendEvent::TerminalResize (11) and TerminalPointer (12), both frontend-gated — the first bump gating in BOTH directions. SUPPORTED=[6..=19].

  • pmacs-protocol/src/terminal.rs now owns the shared terminal bounds, TerminalProcessState, TerminalSelectionSpan, and the single structural policy TerminalFrame::validate; src/terminal/* re-exports them so no duplicate type exists. unicode-width is a workspace dependency so the screen and the validator measure glyph columns identically. MAX_TERMINAL_FRAME_GLYPH_BYTES = 8 MiB bounds the payload instead of widening the transport cap; the measured maximum legal frame encodes to 13,437,863 bytes under the unchanged 16 MiB MAX_FRAME_BYTES. Over-bound snapshots are rejected, never truncated or silently chunked.

  • The Viewport gate keys on the AUTHENTICATED SOURCE'S ACTIVE BUFFER, not the buffer the message names. Viewport also aligns the window to the buffer it declares, so a stale document viewport in flight when a command opens a terminal drags the frontend back off it

    • the terminal then never paints, with no error anywhere. The weaker "is the declared buffer a terminal" reading looks right and fails exactly this way.

  • Suppression compares the COMPLETE ordered payload, never screen_generation: scroll, selection, viewport, and process state all change without advancing it.

  • GPU: pmacs-gpu/src/terminal.rs is a pure cell-space paint planner (testable without a GPU); the renderer builds one shaped buffer per text run so a wide/cluster advance can never choose the next column's origin. pmacs-gpu --headless-probe drives the real attach client without winit (attach::connect_with_sink), which is how criterion 37 gets one real daemon + real PTY + real wgpu path.

  • #135 integrated canonical main after #137 landed. The one code conflict joined TAB_STOP_COLUMNS with the terminal imports in pmacs-gpu/src/main.rs; terminal geometry remains fixed-cell and never consumes the document tab-stop projection.

  • Final post-integration gates: strict Clippy; 1,768 default + 1,944 CRDT library tests; Vterm Stages 1/2/3 at 9/10, 4/4, and 5/7 default/CRDT; statusline 7/8; tab-width 2/2; M4 121; required GPU 139; workspace 2,946 passed across 84 suites; formatting and diff check clean. The first macOS/LuaJIT CI run timed out waiting for VTERM_ALT_READY in the Stage 2 real-TUI smoke; the complete failed-job rerun passed all 12 checks.

  • Stage 2 TUI LANDED ON main — #130 (merge 86fc1bc; docs/vterm-framing.md Revision 7, criteria 15–27). TerminalViewKey keys per-frontend/window projection state over one shared process/screen; logical row anchors retain

    scroll/selection through reflow. One authenticated frontend controls at most one session, with atomic replacement and release on focus/switch/kill/detach.

  • The strict pmacs.terminal Lua surface owns open/state/view/send/terminate and context-implicit scroll/copy commands; the latter error unless the invoking frontend's active window is a terminal. Fixed C-c is the per-frontend terminal escape: only its next key reaches terminal-local editor bindings, while unescaped bound keys pass through to the child. C-c C-c sends one literal interrupt. Copy drains through the acting frontend's clipboard path; active BELs drain once locally and per daemon frontend, while historical/passive bells are baseline-suppressed.

  • TUI composition paints owned terminal cells/styles only inside each window's content rectangle, suppresses document overlays, and keeps sibling splits independent. Daemon key/mouse/paste/focus/resize/detach routing uses the authenticated connection source rather than client-claimed IDs. builtin/runtime/terminal.lua provides the terminal command, view commands, and pure ui.modeline.terminal process/scroll segment.

  • tests/vterm_stage2_acceptance.rs maps Lua transactionality, shared-view isolation, clipboard/modeline behavior, and a hermetic real /bin/sh TUI PTY smoke. Stage 2 changed no wire schema or GPU renderer; protocol remained v18 until Stage 3.

  • PR #130 review round 1 (8702791) aligned dispatch with the approved escape-prefix contract, closed the non-terminal Lua error path, made controller replacement atomic, retained zero-area view anchors, removed duplicate detach work, and replaced per-view deep scrollback clones with borrowed live/published row projections. Focused child-input coverage pins both unescaped bound-key passthrough and C-c C-c.

  • PR #130 review round 2 (b9a7e40) clamps anchors into the first surviving cell when eviction cuts through a wrapped logical line, prevents ambient active_frontend from minting interactive Lua authority, names malformed explicit-context fields, restores dispatcher rationale, and removes owned cell snapshots from terminal mouse routing. The framing now records the transient v18 semantic-controller boundary and bracketed-paste injection deferral.

  • Current-main integration (3f0252f) preserved per-frontend terminal dispatch while applying the landed mode-scoped keymap, and exposed the mode, terminal, and lsp statusline providers together. PR #130 merged at 86fc1bc.

  • Final integrated gate: cargo fmt --check; strict workspace Clippy; 1,753 default + 1,929 CRDT library tests (3 ignored each); mode-system acceptance 1 default + 1 CRDT; Stage 1 acceptance 9 default + 10 CRDT; Stage 2 acceptance 4 default + 4 CRDT; statusline acceptance 7 default + 8 CRDT; M4 114 passed (3 ignored, 1 filtered); required GPU 109; workspace 2,882 passed across 82 suites (19 ignored, 1 filtered); git diff --check clean.

• Tab-width rendering parity LANDED — #137 (merge 2625ec7; docs/tab-width-parity-framing.md rev 2). Source tabs remain one byte while every buffer renderer follows the shared fixed pmacs_protocol::TAB_STOP_COLUMNS = 8.

  • src/display_width.rs owns allocation-free Unicode/tab-aware byte-to-column accounting for plain text, syntax, diagnostics, completion anchors, buffer-style overlays, and search washes.
  • The GPU rich-chunk projection expands source/adornment tabs before cosmic-text shaping and retains first-class source-tab provenance. Carets, hits, selections, peer washes, and diagnostic geometry share the same source/projected boundary rules, including a soft wrap inside one expanded tab.
  • GPU minimap widths use the same tab/Unicode rule and refresh in the accepted text-edit transaction. No config, wire shape, negotiation, or protocol version changed. Local gates: 1,763 default + 1,939 CRDT + 1,763 Lua 5.4 library tests; 2 focused acceptance; M4 121; required GPU 119; workspace 2,911 across 83 suites; strict Clippy and diff check clean.
  • This closes the standing "tab width is a rendering-parity bug, NOT a config gap" deferral in §5: one shared constant now drives every renderer. Terminal cells are deliberately OUTSIDE it — a terminal's columns come from the child, so pmacs-gpu's terminal geometry uses the monospace advance and never TAB_STOP_COLUMNS.

• PARKED: kill-ring browser + persistence. Revision 2 framing is preserved on branch kill-ring-browser, but its 0efb5cd scout is stale and must be repeated before implementation. No PR or implementation is active.

• Roadmap: docs/roadmap-2026-07.md (ranked arcs). Position:

  • Arc 1 (LSP utility surface) COMPLETE — completion popup (#92/#93), panels/references/outline/hover (#94–#96), plus hardening follow-ups (#102, #105, #106).
  • Arc 2 (editing table stakes) COMPLETE — query-replace (#97), kill ring + M-y (#103/#105/#106), comment-toggle (#107), auto-indent (#109), auto-pairing (#110).
  • Arc 3 (persistence) COMPLETE — saveplace/recentf (#98), desktop-save (#99), autosave/crash-recovery (#100), save-clobber fix (#101).
  • Arc 4 (themes + extensibility) COMPLETE — named UI faces (#120), live GPU font preferences (#124), statusline providers (#125).
  • Arc 5 terminal stage COMPLETE — compile mode (#113), Vterm terminal core (#126), TUI frontend (#130), and protocol/GPU Stage 3 (#135) landed.
  • Config registry COMPLETE (#127) — not a numbered arc; it was the cross-cutting substrate ranked first on docs/side-quest-backlog.md's north star, and it unblocks the editing/indent/comment items that were config-blocked.
  • Mode system wiring COMPLETE (#129) — major-mode keymaps, introspection, lifecycle initialization, and statusline display shipped.
  • Locals-query processing COMPLETE — #134 — grammar locals metadata, lexical resolution, settled per-layer facts, shared TUI/GPU local-predicate filtering, and a registry-wide locals-query invariant shipped without a protocol change.
  • Arc 6 (folding) Stages 1 and 2 LANDED — #142 and #149 — the headless fold engine (store, structural source, Lua C-c @ surface, command-path unfold, FoldState production), then the grid/daemon collapse (the VisibleLineMap spine, fold-aware gutter/diagnostics/caret/selection/ presence/viewport/motion, and the interactive unfold widening). Stage 3 (GPU) is next, unframed.
  • Web grammars HTML+CSS LANDED — #146, and LaTeX Stage 1 — #144 with its inline-math parent framing #145.
  • Arc 7 (bottom panel) Stage 1 LANDED — #155 — window placement, window parameters, TUI side windows, the divider, and the adopter display opt-in. Stage 2 (the GPU band) is next and needs its own re-framing; Stage 3 is the default flip. DAP was parked awaiting exactly this arc's Stage 1 and can now re-baseline its touch census.
  • Remaining ranked arcs: 6 folding Stage 3, 7 bottom-panel Stages 2–3, DAP, 8 GPU splits, plus the .ipynb arc (its JSON-grammar prerequisite shipped in #123).

• GPU terminal input LANDED — #166 (main @ b889873; docs/gpu-terminal-input-framing.md rev 2; one review round). The dispatcher applied both terminal-layout syncs to every attached frontend each tick. A semantic session satisfies both conditions — a term_sizes entry from AttachRequest and a terminal declaration — so its PTY was resized twice per tick forever: the grid arm installed the TUI placement size, the semantic arm the declared content rectangle, each arm's old_size == size guard seeing only what the other had just written. The child took a SIGWINCH storm at tick cadence, which made typing into a GPU terminal impossible while output kept flowing. TUI was structurally unaffected.

  • EditorInstance::sync_terminal_layout is split into sync_terminal_controller_liveness (frontend-kind neutral: panel reconcile + release of a controller whose window moved away — reads only views/windows/controller, never a grid size) and sync_terminal_grid_geometry (grid only: TUI placement + resize). sync_terminal_layout survives as the composition, so editor::run and LOCAL are byte-identical.
  • daemon::sync_terminal_layouts_for_tick is the extracted loop body: liveness for every frontend once per tick, then exactly one geometry arm keyed on semantic_states membership — the same fact session establishment uses, so the arms cannot both fire.
  • The trap, kept in a comment: the release on a missing window_placements entry reads like liveness and is grid geometry. A semantic frontend has no placement entry at all, so moving it into the neutral half would release a GPU controller every tick.
  • Why not the one-line guard: the grid arm was also the only per-tick controller-liveness release a semantic frontend got, and sync_semantic_terminal_layout cannot take it over — the buffer-follow snapshot clears the viewport declaration, so that arm stops running in exactly the switch-away case that needs the release.
  • No protocol change (v20). Gates: 1,829 default + 2,006 CRDT library tests; vterm Stage 1/2/3 10/6/9 CRDT; bottom-panel 46; M4 121; required GPU 155; isolated-config workspace sweep 3,177 across 92 suites.
  • Known gap, its own lane: CI never enables crdt, so the Stage 3 real-path acceptance (including a37) is not compiled there. #166's unit pins are not crdt-gated and do run. See docs/active-work.md.

2. How we work (the part that must not drift)

The user is expert and reviews deeply — they falsify framings and find real bugs in round after round. The cadence that has worked for ~40 PRs:

1. Scout ground truth in the code before proposing anything.
2. Write a framing doc — docs/<feature>-framing.md, numbered decisions (Q#XY1…), explicit "Ground truth", "Bets", "Deferred (named)", and an acceptance-test list. Present it and wait for explicit approval ("Go for it" / "Ready to roll"). Expect 1–3 rounds of findings first; revise the doc, don't argue.
3. Branch off main (one feature = one branch = one PR). Commit the framing as the first commit.
4. Implement. Every reviewer finding gets a bite-verified fix — a test that fails without the fix. Watch for vacuously-passing tests.
5. Run the full gate suite (§3). Open the PR with gh.
6. The user replies with "Findings" lists on the PR rounds too. Same discipline. They say when to merge — never merge unprompted.
7. After merge: update this handoff + your memory.

Commit/PR conventions: commit messages via git commit -F <file> (no inline backticks through the shell). Authorship trailers and PR attributions must be truthful: do not add a Claude co-author trailer or Claude Code attribution unless Claude actually contributed. Clippy runs as its own step, never &&-chained.

3. Gate suite (all green before any PR)

cargo fmt --check
cargo clippy --workspace --all-targets -- -D warnings   # own step
cargo test --lib                                        # ~1500
cargo test --lib --features crdt                        # ~1672
cargo test --test <the new/touched acceptance suites>
cargo test --test m4_acceptance -- --skip basedpyright
PMACS_REQUIRE_GPU=1 cargo test -p pmacs-gpu             # 58
cargo test --workspace -- --skip basedpyright           # full sweep
git diff --check

Machine-specific caveats — re-verify on a machine you haven't used before trusting them:

• basedpyright: the DESKTOP's local binary is broken and HANGS the m4_5_basedpyright tests — hence the --skip there. The LAPTOP has a working basedpyright 1.39.9 (verified 2026-07-10: the m4_5 test passes in 0.18s), so the skip is droppable on the laptop.
• GPU on the laptop: AMD Radeon 780M (RADV) — native Vulkan, PMACS_REQUIRE_GPU=1 works without lavapipe.
• Flaky-under-load tests — rerun isolated before treating a sweep failure as a regression. The m8 daemon tests and the m6 process/PTY tests (m6_1_pty_mode_lifecycle_started_then_exited, m6_8_supervisor_reaps_all_children_across_cycles) are timing-based; editor::composition_overhead_under_ten_percent is a render-ratio microbenchmark that fails ~1/3 even isolated single-threaded (already cfg!(macos)-disabled). Vterm Stage 3's merge CI saw one macOS timeout in real_tui_terminal_smoke_restores_host_after_output_input_resize_scroll_copy_and_bell; the complete failed-job rerun passed. The required-GPU gate also failed once in headless_diag_face_recolors_band_counter_despite_unchanged_text, then passed both an isolated single-thread rerun and the full 139-test rerun. A lone timing failure → rerun the test alone (-- --test-threads=1) before investigating. Run the workspace sweep as ONE cargo test invocation piped to a full log — a double invocation + grep -c "test result: ok" can mask real failures with a misleading 0.
• GPU tests need a Vulkan device. PMACS_REQUIRE_GPU=1 makes absence a hard failure instead of a silent skip. Headless option: lavapipe (see docs/repository-audit-2026-07-03.md for the CI harness how-to). On a laptop without discrete GPU, mesa/lavapipe works.
• The desktop's shell is fish (no $(...), use (...); no $UID, use (id -u)). Check $SHELL here before assuming.

4. Substrate invariants (do not undo; tests enforce most of these)

Generated buffers: Buffer::set_generated_contents is the ONE authorized write (terminal copy mode #178) — lift read_only, replace via a single whole-buffer Replace skipping intercepts, discard history, re-assert read_only, and return the Edit. Three things make it a unit rather than a convenience:

• An intercept is not read-only. Buffer::undo reaches the rope through ensure_writable and never consults the intercept chain, so an intercept-only "read-only" buffer is emptied by M-x buffer.undo. Rebinding the undo chords buffer-locally does not close it — compile.lua's own comment says so ("command/menu undo stays dispatchable"). Only rope-level read_only does.
• A bare set_read_only would be worse than nothing, because it also refuses the owner's refresh — the operation such buffers exist for. That is why the pairing, not the setter, is the primitive. There is deliberately no Lua set_read_only.
• A rope write is only half of an edit. The returned Edit must be fanned out (notify_buffer_edit_to_windows, which also queues the daemon-origin CRDT op). Skip it and a displaying window keeps a TextView line index describing the previous contents — the next paint indexes the new rope with stale ranges and trips assertion failed: end <= self.len() — while replica mirrors never import the write at all.

History clearing is load-bearing twice (nothing can pop entries read_only makes unreachable, so they leak), and must clear whichever history the buffer has: the v0.1 stacks are bypassed in CRDT mode, where it lives in loro's UndoManager. That has no clear, and needs none — a manager records only what happens after construction, so CrdtState::clear_undo_history rebinds a fresh one to the same doc.

Not yet adopted — the inventory is four writer mechanisms covering five buffers. Every remaining intercept-protected writer uses the older idiom: an erroring intercept plus set_round_trip_input, written through bypass_intercept, with the rope left writable. All are emptiable by M-x buffer.undo:

writer	buffers	shape
builtin/runtime/listview.lua:60-61	every listview panel	delete-all + insert
builtin/runtime/compile.lua (ensure_slot)	*compilation*, *shell-command*	append per output batch
builtin/commands/default.lua:869	*search-results*	reset per query, then append per match batch
builtin/runtime/dired.lua:371	every dired buffer	whole-buffer replace

Do not read ensure_slot as covering the search panel — it serves *compilation* and *shell-command* only (compile.lua:1090,1125). *search-results* is an independent panel with its own intercept, round-trip mark and writes, and compile.lua names it only in a predicate. Nor is the scope "every generated buffer": *workers*, *help* and *buffer-list* are generated too but do not use this idiom, and the REPL package's intercept (builtin/packages/repl/init.lua:187) is an op-filtering editing policy, not a read-only panel — neither group belongs to this lane.

Adoption is not a one-line swap. It inherits the fan-out obligation, and the three appending buffers need a streaming variant of the primitive; listview and dired already write whole-buffer replaces and are the cheap half. Recorded in COHERENCE.md §14.

And it does not replace set_round_trip_input. The protection is layered across two copies: rope-level read_only refuses the op at the daemon; round-trip input stops a semantic frontend applying optimistically to its own mirror, which a daemon-side refusal cannot reach — the refusal arrives after the frontend has already painted, so it buys divergence, not prevention.

Command boundaries (Arc 2 kill-ring substrate) — EditorCore.command_history: HashMap<FrontendId, CommandBoundary{this, last}>, per frontend. Rotate on: keybound command, self-insert, menu invoke, invoke_interactive (the M-x path). Break on: unbound key, GPU optimistic CRDT edits, pointer gestures (wheel scroll deliberately does NOT break), unified paste. Plain pmacs.command.invoke stamps nothing (programmatic API); invoke_interactive rotates-then-invokes (Emacs M-x semantics). Single-codepoint optimistic CRDT inserts classify as buffer.self-insert (exact decode; "a(" breaks instead). Lua: ed.this_command() / ed.last_command().

Effective-edit returns — buf:insert/delete/replace return the post-intercept (start, end, inserted_len). Callers that care (killring, comment) compare EXACTLY against the request; length-delta and text-at-position checks are documented defeated patterns. Always pcall the mutator: a rejecting intercept must report, not throw through, and failed ops must leave no state (kill chains, yank sessions).

Kill ring — entries {id, text} with stable monotonic ids; chains and yank sessions are per-frontend and id-checked (an index is not stable under other frontends' pushes). OS clipboard mirrors to the acting frontend only. Paste is a unified daemon arm keyed by the dispatcher's AUTHENTICATED source — never a payload frontend_id.

LSP outbound positions — every Position/Range builder in src/lsp.rs routes through outbound_position (byte → negotiated encoding). Any new request builder must too; UTF-16 servers reject raw byte columns on non-ASCII text. Semantic tokens: full, full.delta, and range are three INDEPENDENT capabilities — gate each.

Persistence (Arc 3) — state-dir wiring lives in install_state_dirs() on real entry points only, NOT EditorState::new() (tests stay hermetic); PMACS_STATE_HOME overrides. Autosave: one buffer owns a path's recovery slot; only recover/discard release unclaimed crash data; adopt clears the old owner's skip cache.

Protocol — encoding-breaking bumps are deliberate and versioned. Canonical main is [6..=20]. The in-review bottom-panel 2B-1 schema extends support to [6..=21], while ADVERTISED_PROTOCOL_VERSION stays 20 until 2B-3 provides compatibility-preserving activation; the server-first Hello cannot advertise 21 without stranding existing v20 clients before AttachRequest. v15 = CompletionPopup + StatusFacts.message; v16 = ThemeFacts; v17 = FontFacts; v18 = StatuslineSegments; v19 = the vterm terminal family; v20 = semantic SessionBootstrapRequest plus appended InitialTargetResult; v21 reserves the panel frame/event family. New wire surface ⇒ bump + both-frontends support + acceptance. An APPENDED variant must be guarded by a byte pin on the PREVIOUS final variant — its own round-trip cannot detect a discriminant shift.

Fake LSP (src/bin/pmacs_fake_lsp.rs) modes: fullonly, rangeonly, rangeonly16 (UTF-16 + fail-closed bounds validation), sighelp. Use these for capability-matrix tests, not real servers.

5. Hard-won ops lessons

• A test that skips on a missing precondition reports ok, and a gate log cannot tell that apart from a pass. vterm_stage3_acceptance::a37 — the only acceptance driving a real daemon, a real PTY and a real wgpu render together — derives pmacs-gpu from CARGO_BIN_EXE_pmacs and, when that binary is absent from the target directory, prints a skip and returns. A fresh worktree reports the suite 9/9 in 0.17 s having never run it; a real run takes ~4 s. PMACS_REQUIRE_GPU=1 is what promotes the skip to a failure, and the standing gate list applies that flag to cargo test -p pmacs-gpu, a different package. Two habits follow: build the workspace before believing any suite that reaches for a sibling binary, and judge such a suite by its elapsed time, not its verdict.

• Before attributing a red test to your branch, run it on the merge base. a37 failed on the #173 branch, which looked like a regression; it failed identically on the PR's own base and on two intermediate commits, and had passed on that same base twenty minutes earlier. The variable was machine load from a second agent compiling continuously. Load-sensitive tests make both verdicts uninformative in isolation, so the base-commit run is the cheapest way to tell a regression from weather — and it is much cheaper than the bisect it replaces.

• A daemon-side fix is not deployed until the daemon is restarted from a tree that contains it. #166's reporter rebuilt and saw no change: the running daemon had been started from a shared checkout still on a pre-fix branch, and pmacs --gpu attaches to whatever process already owns the socket. Rebuilding a binary does nothing to a running process. When validating a daemon-side fix by hand, check the running process's binary path and start time against the tree you think you fixed — ps -eo pid,lstart,args | grep '[p]macs --daemon' — before concluding the fix failed.

• Two operations that must be alternatives are not made alternatives by being adjacent. The dispatcher applied its grid and semantic terminal-layout syncs to every attached frontend; a semantic session satisfies both conditions, so its PTY was resized twice per tick forever and the child took a SIGWINCH storm that made a GPU terminal untypable while output still flowed. Each arm had a correct old_size == size idempotence guard — individually sound, jointly useless, because each saw only the size the other had just written. Write mutually exclusive per-frontend-kind work as one if/else keyed on the same fact session establishment uses, and extract the loop body so a test can drive the real thing.

• Bite against every pre-image the fix could plausibly have taken, not just main. For the same defect, the obvious one-line guard (skip the grid arm for semantic frontends) does fix the storm — and silently introduces a controller leak, because that arm was also the only per-tick controller-liveness release a semantic frontend got. A single revert would have scored the fix complete. The pin that catches it (acc 6) deliberately passes on main and fails only against the naive guard: today's defect supplies the release by the accident of running an arm it should not.

• A quiet child is an instrument. A frame storm is invisible against a fixture that legitimately emits hundreds of frames, and an assertion like frames >= 2 cannot see one. The same applies to geometry: a "did a frame at the new width arrive" readout is satisfied by a geometry oscillating through that width. Assert upper bounds over a fixed window against a child that produces nothing, and let the child self-report the signal you care about (a SIGWINCH trap printing a fresh distinct breadcrumb per signal — repeated identical markers paint nothing, because cell::diff skips already-matching cells).

• TerminalMode::Raw makes sh-based input fixtures useless. There is no ICRNL, so Enter delivers CR and a read -r loop waits forever for a LF that never comes — the test then "proves" input never arrived. Use exec cat, which copies stdin to stdout byte by byte. It is also the right echo instrument for the opposite reason people assume: termios ECHO is off in raw mode, so nothing double-echoes and one keystroke yields exactly one cell.

• Draining the event stream to learn a pid also TICKS, and a tick reaps. #176's observing_the_leader_does_not_consume_the_exit_event failed under load with "process ProcessId(26) is not running": its helper drained for Started to read the pid, drain_until ticks while it drains, and a tick can observe an immediately-exiting child and move the record out of Running — after which signal never reaches the code under test at all. It passed standalone only because the drain returned on Started before poll_one saw the exit. Read the pid straight from the supervisor record, which does not tick, and fail fast if the record has left Running. Verified load-bearing under matched load: 0/15 failures with all 16 cores saturated versus 1/10 for the ticking helper. The general rule: an observation helper that advances the system is not an observation.

• A wait predicate that is WEAKER than the assertion it guards is a race, on every platform that happens to lose it. #174: the m4 config sink test waited for contains("probe") and then asserted "probe":true — six bytes further on. The sink is JSONL written by a separate process, so the test could read a half-written line; Linux won that race reliably and macOS/lua54 did not, reporting the truncated {"rust":{"probe":. Any "wait until the file mentions X, then assert Y about the file" shape races whenever Y is stricter than X. The fix is to wait for the exact unit the assertion reads — here a trailing newline, true only once a whole writeln! record has landed, and still correct if the payload's field order or spelling changes. Two things generalize beyond the fix:

  • A race you cannot reproduce can still be bitten at one remove. The failure needs a scheduler outcome Linux does not produce, so no local run falsifies it. But pump_async asserts on its deadline, so replacing the predicate with an unsatisfiable one proves the wait is load-bearing rather than decorative — and restoring the old predicate, which still passes locally, proves a green local run cannot distinguish the two. Bite what you can reach and say plainly which claim rests on structure instead of evidence.
  • The obvious fix is sometimes worse. The sibling m4_26 has the same weak shape and was deliberately left alone: waiting for the expected value would convert a genuine rootUri regression — what its assert_ne!s exist to catch — into a five-second timeout with a misleading "server didn't initialize?" message, trading a precise assertion diff for a vague hang. Closing it properly needs a record terminator in the fake server first.

• Proving a child has exited is harder than it looks on this codebase. A fixed sleep is not proof; nix's waitid is unavailable on macOS; and libc::waitid needs unsafe, which #![forbid(unsafe_code)] rules out. What works is driving the production diagnostic in a bounded loop until it observes the exit. Relatedly, #176's round-1 review replaced every substring assertion with exact message equality built from the kernel-assigned pid — the substring forms would have accepted a hardcoded target or a wrong exit code.

• The checkout may be shared with the user. Check git status for foreign uncommitted work before any stash/checkout/branch surgery; never assume dirty files are yours. (Their uncommitted fix was nearly orphaned once.) A clean status goes stale within minutes when two lanes are active — for parallel work, git worktree add a sibling directory off main instead of switching the shared checkout.

• Never git stash in this repo. The stash namespace is REPO-GLOBAL — one list shared across every worktree and with the user; a scripted push/pop can pop a human's years-old stash into your tree (happened during #111: a failed stash push chained into stash pop, which grabbed the user's PR-#17-era entry). For run-tests-against-an-old-version swaps, use scripts/bite — a trap-guarded one-file swap over read-only git show, with an inverted verdict (exit 0 iff the tests FAIL against the old version), making bite-verification machine-checkable.

• A fix must be COMMITTED before it is bitten. scripts/bite restores by git checkout --, which reverts the file to HEAD, not to the state it found — so any uncommitted work in a bitten file is destroyed. A whole review round's fixes were wiped this way during #165. Corollary for a NEW file: the swap-over-git show mode does not apply at all, so its claims must be bitten by hand-editing, which makes the commit-first rule load-bearing rather than hygienic.

• A CONFLICTING PR silently runs no CI at all. GitHub builds pull_request workflow runs against the PR's merge ref, which it does not create while the branch conflicts with its base. So pushes land, the branch updates, no run is ever queued, and nothing reports the absence — the checks list simply keeps showing the last successful run, which reads as current. Three pushes to #165 produced zero CI before the cause was found, and gh pr checks returns nothing usable here. On any lane that lives through a moving main, check gh pr view <N> --json mergeable,mergeStateStatus,headRefOid and confirm a run exists for the current head sha, not merely that a recent run was green.

• Stacked PRs: retarget the child to main BEFORE merging the parent — GitHub auto-closes a PR whose base branch is deleted and cannot reopen it (#104 → re-opened as #105).

• Frozen reviewed PRs do not absorb moving overlapping work. For #135 and #137, the approved/frozen #137 landed first; the larger #135 lane then merged canonical main, retained its review anchors, and reran every gate. Derive that integration surface from git diff <base>..main, not the other PR's file list — concurrent landed work added an overlap the original two-PR comparison missed.

• Scripted edits (sed/python) in files with repeated similar blocks (src/lsp.rs JSON builders): anchor on a unique line or you will silently edit the wrong block. This produced a vacuously-passing test and cost an hour.

• Acceptance fixtures that open .rs/.py files must empty pmacs.lsp.config first — the after-load hook spawns real servers (rust/python/c have default configs). Language detection (grammars + pmacs.lsp.filetypes) is unaffected.

• Test scratch buffers have no path ⇒ no language; use tempdir files when language matters.

• Dual-purpose session state is a reset-contract trap (#120 rounds 2–5): last_status doubled as the peer emission baseline AND the stale-diag count freeze, so resetting baselines on BufferSnapshot zeroed mid-edit counts. Knowledge about a buffer belongs in shared stores (DiagnosticStore severity totals), never in per-session baselines; and any daemon-side reset needs its frontend mirror audited in the same round (the GPU snapshot arm missed search/menu/status the first time).

• Tab width is a rendering semantic, NOT a config gap. The implementation on tab-width-parity fixes the width at the TUI's established 8 columns, shares that constant through pmacs-protocol, and expands tabs only in each display projection. Defining editor.tab-width could not have fixed the GPU: source text and semantic spans stay byte-addressed while cosmic-text needs projected spaces plus an inverse hit/caret map. A future configurable width would require a buffer-effective frontend fact and cache invalidation; do not re-plan it as a scalar config-only change.

• A test that never runs passes. Two #127 review-round tests passed vacuously at first: pmacs.editor.save() is the RAW save, while buffer.before-save fires inside the buffer.save COMMAND (builtin/commands/default.lua), and save() no-ops on an unmodified buffer — so a fixture that opens a file and saves it asserts on bytes nothing rewrote. Dirty the buffer with a real edit and go through pmacs.command.invoke("buffer.save"). Caught only because the other case failed and the cause was chased instead of the assertion adjusted.

• A message that ALIGNS state cannot be gated on the state it names. FrontendEvent::Viewport both declares a byte range and switches the frontend's window to the buffer it names. Gating the vterm v19 dual declaration on "is the DECLARED buffer a terminal" therefore left a stale in-flight document viewport free to drag a frontend straight back off a terminal a command had just opened — the window oscillated, the terminal declaration was refused every time, and no frame ever arrived. Nothing errored. The gate has to key on the authenticated source's ACTIVE buffer. Generalizes: when two messages declare competing views of "what am I showing", the arbiter is the daemon's own state, never the claim inside either message.

• A pass that sets a mode flag must clear it on EVERY exit. The semantic producer's terminal pass returned early via ? when no declaration existed, leaving terminal_active set — and the daemon uses that flag to suppress CursorByte and the presence sweep, so a frontend that went back to a document silently lost both. Caught by an acceptance assertion, not by any type.

• Sub-crate acceptance needs a real seam, not a fixture. pmacs-gpu depends only on pmacs-protocol, so "real daemon + real PTY + real wgpu in one path" could not be an in-crate test. Generalizing attach::connect's reader sink (connect_with_sink) and adding --headless-probe gave the acceptance the REAL handshake, outbox, writer, and render_to_view — which is the whole point; a decoded-message fixture would have proved none of the three fit together. The probe found two real defects the in-process tests did not.

• Real-grid acceptance must budget for macOS startup and path width. A 100 ms first-Hello timeout failed under loaded macOS CI; use the normal five-second handshake window, then short polling reads. An 80-column split also clipped a custom statusline segment after macOS's long /var/folders/... temp path while passing on Linux; size the grid for the longest supported fixture path (the mode-system test uses 160 columns per split).

• A provisional session that fails mid-bootstrap must actually close its socket, not just drop its handle. #148's dispatcher-side target-failure paths wrote InitialTargetResult::Failed and dropped the write-half UnixStream clone, but a clone shares the underlying FD — the per-attach reader thread stayed alive with no installed session state. A client that kept the socket open past Failed and sent any ordinary event hit an .expect reachable only through the new failure path and panicked the whole daemon. Fix: shutdown(Shutdown::Both) on every dispatcher-side failure path, plus a defense-in-depth session-registry membership check before any FrontendEvent touches render/size/editor state. Generalizes: when a new failure path can leave a handle installed without its owning session, dropping a handle is not the same as tearing down the connection.

• A guard with no production caller passes every direct-call test. #155 round 1: EditorCore::try_split_active implemented the side-window split refusal, but pmacs.window.split_horizontal / split_vertical — and therefore C-x 2 / C-x 3 — still called plain split_active. The acceptance test called the core method directly, so reverting the guard entirely would have left every test green. Same shape as folding #142 round 2. Assert through the outermost user-reachable seam (try_exec(&s, "pmacs.window.split_horizontal()")), then falsify by revert. When the test shares a file with the code it pins, scripts/bite cannot swap it — break the production line by hand and git checkout --.

• A geometric readout is not a state predicate. TerminalViewStatus::at_bottom is defined as scroll_offset == 0 — "the viewport currently reaches the tail", not "this view follows the tail". A still-anchored view satisfies it whenever it happens to be tall enough, so asserting it could not detect that Q#BP7's growth re-arm had never been implemented (#155 round 2): the next rows the child printed pushed the anchored view back into history. Pinning following requires advancing the world — feed more child output through a filesystem gate — and asserting the view came along. Related: scroll_offset is viewport-relative, so "unchanged across a height change" is vacuous or wrong; the invariant is the frozen ANCHOR.

• A PTY in the default mode does not translate LF to CRLF. An echo-driven test fixture staircases rightward, and past the viewport width every row clips to blanks — so assert_eq!(top_before, top_after) compares "" == "" and passes for any regression (#155 round 2). Emit printf '...\r\n', and guard text comparisons with assert!(!observed.is_empty()) the same way the panel daemon pin guards on !panel_hidden.

• Widening an ambient resolver into a scoped one can make a total function partial. #155 round 2 resolved both arms of pmacs.window.buffer() through the acting frontend "for uniformity". acting_frontend follows the interactive origin, which can name a frontend with no registered view (a bare dispatch_key from an unattached peer), so the no-argument arm began raising instead of answering. No runtime caller pcalls it, so killring, syntax, autosave, pair, indent and comment silently dropped operations — kill_ring_acceptance went 30/30 to 25/5 on every CI platform. The ambient resolver's fallback is what makes it total; keep it, and document that as deliberate. Uniformity is not free when the paths have different totality.

• An upgrade decision must be tracked independently of the outcome that triggered it. #148 published a target's fresh BufferSnapshot to existing grid replicas only when the buffer was newly_loaded || newly_created — but a target can dedup onto an already-existing, not-yet-CRDT-backed buffer (e.g. one a startup hook created via find_or_open and never activated), silently upgrading it without telling pre-attached replicas. The later F29 lazy-upgrade sweep then saw an already-backed buffer and never broadcast it, permanently stranding those replicas on v0.1 round-trip for that buffer. Fix: have the upgrade helper report whether it performed the upgrade, and OR that into the publish decision rather than inferring it from the caller's own load/create branch.

6. Named deferrals (the standing backlog, consolidated)

Editing: word kills (M-d/M-BS — need bytes-returning deleters + prepend-on-backward append), C-SPC set-mark, C-u C-y / C-M-w, kill-ring browser + persistence, clipboard watching, block comments + mid-line comment spans, comment-dwim append-at-EOL, per-language comment padding. Pairing (framing "Deferred"): wrap-region on opener, pair-aware backspace, RET-inside-pair closer-on-own-line, in-string/in-comment inhibit (needs node-at-byte pmacs.parse), undo amalgamation (pair = one step), balance-aware quotes (the per-buffer toggle SHIPPED in #127 as editing.auto-pair). Editops deferrals (full list in its framing): recenter (blocked on viewport facts — the GPU never consumes daemon view_top), Unicode case/word classes, region-spanning move/duplicate, locale collation for sort-lines, ensure-final-newline on save. Substrate: buffer-aware edit epoch (after-edit currently compares the ACTIVE buffer only), wire provenance for CRDT self-insert classification, Lua intercept probe, completion.lua still on the old cursor-delta heuristic (migrate to this_command), the TUI's nonempty-selection optimistic type-over gate, generated-buffer search invalidation, cross-peer chronological undo arbitration (mixed source/daemon history; pinned by auto-pairing acceptance), origin-pinned buffer.after-edit fan-out (a context-switching intercept changes what later callbacks — LSP, completion — observe). LSP/persistence: hidden-buffer LSP attach, daemon desktop-restore, the warning half of external-change detection (verify-visited-file- modtime). Config registry (SHIPPED #127; these are its own named deferrals): persistence of settings and the custom-file split-brain question, M-x list-settings as a listview panel, a settings completion source for the minibuffer (minibuffer.read's source is a fixed Rust-side vocabulary), table-valued settings (so pmacs.lsp.config, pmacs.pair.sets, pmacs.comment.strings and the pmacs.parse.* write-through proxies stay raw Lua), migrating the remaining scalar setters (async_config ×2, killring.max, the enable booleans) and pmacs.gpu.set_font, pending-set staging for names defined after init.lua runs, and a scope = "global" define flag — set_local is currently accepted for autosave.interval-ms, where a per-buffer value is meaningless. Tab width is NOT a config gap — see §5. Mode system (SHIPPED #129): minor modes, buffer.after-mode-change, mode-scoped settings, describe-mode, and persistence of explicit major-mode overrides/clears across sessions. Modeline detection (SHIPPED #132): bounded first/last-line Emacs -*- and Vim ft=/filetype= parsing, explicit-over-inferred precedence, alias normalization, and shared fresh-load language pinning for syntax/highlight/LSP startup. Highlight/detection (from the #114–#118 side-quest + injections #122): locals-query processing SHIPPED #134; remaining injection follow-ups now that the engine landed (#122) — injection.combined (many matches → one shared parse; PHP-in-HTML, some comment schemes), child-tree incrementality + range-scoped layer rebuild (child layers cold-reparse on every settle today), injectable runtime/Lua-registered languages (v1 resolves only against BUILTIN_LANGUAGES), and the next injection consumers gated on new grammars — HTML/CSS/GraphQL/SQL (<script>/<style>, JS/TS template literals, doc-comment code); modeline detection as a 5th layer (-*- mode: … -*- / # vim: ft=…) SHIPPED #132; byte-accurate multibyte cursor placement in move_active_cursor_to (still steps one codepoint per LSP byte column). A full Jupyter .ipynb setup (reader → editable → kernel execution) now has its JSON grammar prerequisite, but remains a real arc, not a one-shot. GPU: auto-reconnect after daemon restart, splits/multi-buffer, gutter riders (whitespace guides, folding, git markers). Themes (full list in theme-faces framing rev 9 "Deferred (named)"): popup/menu/dropdown bg + selected-row faces, ui.background / ui.caret, ui.modeline.inactive, minimap chrome, peer-cursor palette (+ui.selection for peer rects), ui.inlay_hint (needs the epoch treatment on its producer), wire alpha, Indexed palette unification, named-theme registry / light theme / persistence (the registry exists now, #127; theme persistence still waits on settings persistence), grid-vs-wire default_style asymmetry, mask widening (gutter bg, wash glyph recolor, statusline bg echo surface, chrome bold/italic/underline re-shaping). Housekeeping: F-016 lua_bindings/mod.rs split paused mid-way (tranches 0–2 landed, ~5–8 PRs left; see docs/lua-bindings-split-framing.md). Full cross-cutting index of the non-themes backlog (this list + every framing doc's Deferred section + a code sweep, themes excluded, with a prioritization north star): docs/side-quest-backlog.md.

7. Machine-local facts (desktop) that do NOT travel

Three untracked files live only on the desktop working tree and are deliberately never committed: docs/pmacs-gpu-editing-perf-handoff.md, docs/session-5-stale-styling-handover.md, python_experiment.md. Don't expect them in a clone; on the desktop, never delete them.

8. Update protocol for this file

When a PR merges, an arc opens/closes, or a decision lands: edit the snapshot (§1), append lessons (§5) and deferrals (§6) as they arise, bump the date line at the top, and commit — usually riding the same PR as the work. Keep durable architecture here; put branch hashes, machine-local tools, incomplete verification, and recovery commands in docs/active-work.md. This is a briefing, not a log: prune sections that stop being true.