neuwirth
/
pmacs
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
pmacs/docs/active-work.md

56 KiB
Raw Blame History

Active work — cross-machine resume ledger

Snapshot: 2026-07-26. This file records volatile work that has not landed on main. Read it after docs/agent-handoff.md. Remove completed entries when their PR merges; do not let this become a second permanent backlog.

One lane below is retained past its merge, and says so at its own head: the PTY terminate diagnostic (#176), because no landed-doc PR owns moving its facts to docs/agent-handoff.md yet, and rule 4 removes a lane only after that move. Every other merged lane has been removed — the Lean 4 and GPU-terminal-input headers this paragraph used to disclaim are gone, as are the inline-math (#172), dired (#169), and terminal config + copy mode lanes — the last of these was #180's work, folded into #182 so two open PRs would stop re-conflicting in this file.

Trust the canonical-base line below over any lane header: if a PR number appears in git log --first-parent githubsucks/main, it has landed regardless of what a lane says.

Repository authority

  • Canonical development URL: https://github.com/levineuwirth/pmacs.git. This ledger uses the normalized local alias githubsucks so its refs and recovery commands are identical on every machine. Remote names are otherwise machine-local: origin may name this canonical URL, a release mirror, or something else, and therefore has no authority by name alone.
  • Canonical base at this snapshot: githubsucks/main @ 42025e4 (Lean 4 Stage 4b #181, atop 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, and the minimap blank-slab fix #159; protocol v20). The previous snapshot named 74301d1, and the recovery floor advances with it: the check below now requires 42025e4 or newer, so a tree at 74301d1 no longer passes. That is deliberate — the floor moves with the base, because a check that accepts an older commit than the declared base passes on a tree the rest of this file does not describe. Lanes below that name an older base have not been re-based; derive their integration surface from git diff <their base>..main.
  • On the transfer source, origin/main named a release mirror at d3fa632 and lagged badly. On the current destination, origin names the canonical URL. This difference is why all recovery begins by verifying URLs and normalizing githubsucks rather than trusting origin/main.
  • The shared desktop checkout contained unrelated uncommitted work. The branches below were prepared in isolated worktrees; never clean or overwrite the shared checkout to recover them.

Start on another machine by inspecting its remotes:

git remote -v
git remote get-url githubsucks

If the second command says the alias is absent, add it; if it prints a different URL, stop and resolve that collision rather than overwriting an unknown remote:

git remote add githubsucks https://github.com/levineuwirth/pmacs.git

Then recover current refs:

git fetch githubsucks --prune
git log -1 --oneline githubsucks/main
git worktree list
git status --short --branch

The git log command must expose 42025e4 — the base named above — or a newer intentional main. Keep this threshold and the canonical-base line in step: a recovery check that accepts an older commit than the base it declares canonical will pass on a tree the rest of this file does not describe. If it does not, stop and repair the remote/fetch configuration.

PTY terminate diagnostic lane — MERGED (PR #176)

Lane retained deliberately, and it is the next one to close. #176 merged into main @ bf8878f (2026-07-26); rule 4 below removes a merged lane, but only after its durable facts reach docs/agent-handoff.md. That absorption is unowned — no landed-doc PR exists for #176 — so removing the lane now would delete the record instead of moving it. Whoever opens that PR removes this section.

  • Portable branch: githubsucks/pty-terminate-eperm; worktree ../pmacs-math-slice. PR #176, base main, based on ccf29e3 with c93f9ee (#175) merged in.
  • Approved framing: docs/process-signal-tolerance-framing.md revision 4, after three review rounds.
  • Diagnostic only. No disposition change. 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.
  • Why nothing is fixed: revisions 1–3 each proposed a tolerance rule and all three were rejected as unsound in the same way — each concluded something about a process from something that was not about that process. Rev 1 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.
  • Two facts that killed the original argument. group = true is rejected for PTY mode at spawn (src/process.rs:1428-1429), so the reap ledger never applies to the PTY path at all; and the ledger comment (:1075) says EPERM "cannot happen for our own children" and drops the entry for bounded growth — not a ruling that EPERM means dead.
  • The CI evidence never established the child had exited. The probe's last source statement is a file write and CPython teardown does not synchronise with it, so no tolerance rule could even be shown to fix the symptom. That is the whole reason the lane is diagnostic.
  • 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.
  • Not "strictly additive". try_wait reaps and caches, so an exited child may be reaped earlier than otherwise. Safe 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.
  • Round-1 review fixes: the exited-child tests no longer use a fixed sleep as proof of exit (nix's waitid is unavailable on macOS and libc::waitid needs unsafe, which the crate forbids), instead driving the production diagnostic in a bounded loop until it observes the exit; and every assertion is now exact message equality built from the kernel-assigned pid, since the substring forms would have accepted a hardcoded target or a wrong exit code.
  • Bites, all verified rather than assumed: tolerating the failure fails the disposition test; stubbing the leader observation fails three tests including the one-event pin; a hardcoded target fails four; a wrong exit code fails two.
  • The sweep found a real defect in these tests, not a flake. observing_the_leader_does_not_consume_the_exit_event failed with "process ProcessId(26) is not running": the pid helper drained for Started, and drain_until ticks. A tick can observe an immediately-exiting child and move the record out of Running, after which signal never reaches the diagnostic at all, so the bounded loop spun to its limit. It passed standalone because the drain returned on Started before poll_one saw the exit; only load lost the race. Fast-exiting children now read the pid straight from the supervisor record (no tick), and the loop fails fast if the record left Running. Verified under matched load: 0/15 with all 16 cores saturated, while the old ticking helper fails 1/10 — the fix is load-bearing.
  • Verification: fmt, git diff --check, strict workspace clippy clean; lib 1,838 + CRDT 2,015 (both +6, exactly the new tests); GPU 202; M4 121; bottom-panel 46; compile-mode 67; vterm 9/6/5; isolated-config --no-fail-fast sweep 3,258 across 93 suites, zero failures. Earlier sweeps on this branch showed two failures and then one; the totals reconcile (3,256/2 → 3,257/1 → 3,258/0, same test count). The two that were genuinely unrelated — read_dir_supersede_cancels_in_flight_predecessor (known pre-existing) and headless_snapshot_round_trip_summary_restores_the_minimap — are load-contention flakes; the second is structurally unreachable from this diff, since pmacs-gpu depends on pmacs-protocol and never on pmacs.
  • Parked, each with its reason: all tolerance rules (need the evidence this PR produces); terminate idempotence for an already-reaped process (independent fix, different failure, one feature per PR); and signal_target's read-then-kill of tcgetpgrp — still the most likely real fix site.
  • The lane closes when this merges. It does not wait for the flake to recur; the next occurrence carries its own evidence under whoever's PR, and a Stage B framing follows then.

Lean 4 lane (Arc 8) — Stages 1–4a MERGED; Stage 4b IN REVIEW

  • Stages 1, 2, 3a, 3b and 4a are MERGED — #160 (main @ 0827dd1), #161 (46a1b8f), #167 (6f348c9), #170 (d400f30), #179 (a27f646). Their full histories were pruned from this ledger in round 6, per this file's own instruction to remove entries when their PR merges; the durable facts now live in docs/agent-handoff.md §1's Lean 4 bullet, which is where a fresh machine should read them. docs/lean4-mode-framing.md rev 9 carries the decisions.

Stage 4b — the Unicode input method (branch lean4-stage4b-input-method)

  • Framing docs/lean4-mode-framing.md revision 12 (rounds 10, 11 and 12 = review of the implementation). Stage 4a (the typed-edit consumer chain) MERGED as #179; this branch is 4b, the Lean content that registers on it.

  • Footprint: scripts/regen-lean-abbrev (new, the generator), builtin/runtime/lean_abbrev.lua (new, VENDORED DATA — 1,855 entries from leanprover/vscode-lean4@17d1d08, Apache-2.0), builtin/runtime/lean_input.lua (new, the consumer at priority 50), src/editor.rs (two include_str! blocks), tests/lean_input_acceptance.rs (new, 31 tests), and one #[cfg(feature = "crdt")] --lib test in src/daemon.rs (acceptance 45f). No protocol change (Q#LN14). Entirely Lua apart from the load sites and that one test.

  • Round 9 corrected three acceptance criteria that the real table contradicts — found by simulating the state machine over all 1,855 entries and re-reading upstream at the pinned commit, not by reading the prose again. \to is NOT eager (top, to0, toa extend it); \zzzz expands to ζzzz because ze/zeta/zsqrtd exist, and only $ % , ; @ W open no key at all; and \alpha's undo does not restore \alpha because alpha IS eager, so the terminator is a separate edit. Criteria 38, 41 and 42 now state both paths.

  • Two generator bugs, both caught by its own round-trip check failing closed: str.splitlines() also splits on U+2028/U+2029, and 53 symbols contain one literally, so the check reported a count mismatch that was its own bug; then escaping via chr(byte) produced a latin-1-shaped string that write_text(encoding="utf-8") re-encoded, and every non-ASCII symbol landed double-encoded. The first version of the check compared IN-MEMORY strings and agreed with itself. It now stages the file, re-reads the bytes from disk, and renames into place only on a match.

  • The point must be placed explicitly after the replace. The expansion SHRINKS the buffer (\alpha 6 bytes → α 2), so a point left at the pre-edit offset is past the new end and every later self-insert is silently rejected — the editor looks dead after the first expansion. Pairing's "no cursor motion on the clean path" does not generalize: that holds only for an insert AT the cursor.

  • Three tests were vacuous when first written and were found by biting, not by review: the abandonment test asserted text that a wrongly-surviving record would also produce (claiming makes no edit — it needed the follow-up keystroke that completes an eager key); the re-arm test used the framing's own \alpha\to, which never reaches the re-arm branch because alpha is eager and closes the record first (\al\to does); and both buffer-switch tests passed through find_or_open's fresh-load path, which fires buffer.after-load and a record-less edit rather than buffer.after-switch — deleting the subscriber left them green. All three now bite.

  • Bite table (each mutation, and the tests it fails):

    Mutation Tests it fails
    register at priority 150 (after pairing) 2
    claim only completed expansions 2
    longest match instead of shortest 9
    equal-length tie keeps the LATER key 3
    remove the eager branch 8
    expand without the terminator in the span 2
    remove the re-arm branch 1
    remove the point-still-at-span-end check 1
    remove the exact-revision check 1
    leave the point where the replace found it 5
    remove the lean4 language gate 1
    remove the lean.abbrev gate 2
    buffer.after-switch clears every frontend 1
    delete the buffer.after-switch subscriber 1
    frontend.detached purges every frontend 1
    claim the terminator 1
    expand inside the chain, then decline 2
    drop the cursor() == post_cursor check 1
    place the point without the context guard 1
    load lean_input.lua after lsp.lua 1
    let a nested fan-out consume the deferred slot 1
    stop counting chain invocations 1
    count fan-outs in the expander instead of the sentinel 1

    Acceptance 45f bit by construction: without a registered window for the source frontend it ran six fan-outs with a nil record and proved nothing, because handle_remote_crdt_op arms nothing unless the source's active window displays the buffer.

  • Round 10 (review) found three defects, all about what happens AROUND the expansion rather than about resolving an abbreviation. A pair character that TERMINATES an abbreviation never reached pairing (\alp( gave α(): the first revision claimed the terminator, and merely declining is not enough either, because the chain hands each consumer a copy of the record made before any consumer ran — so expanding inside the chain invalidates pairing's copy and the closer is lost anyway (verified by mutation, not assumed). The expansion now runs on its own buffer.after-edit subscriber after the chain, with a span that stops before the terminator. That is a new instance of Q#AP7, so it is now pinned with the sighelp fake server. Post-insert point motion was also mistaken for a valid span (the relevance check needs cursor() == post_cursor, as pairing's has since #110), and cursor placement could move a buffer an intercept had switched to.

  • Round 11 found the round-10 fix incomplete in one place: buffer.after-edit fan-outs NEST. A consumer between the expander (50) and pairing (100) that calls pmacs.hook.run("buffer.after-edit") re-enters the expander's subscriber while the OUTER chain is still mid-list; the nested pass expanded and outer pairing then resumed with an invalidated record — α( again, through the chain's documented re-entrancy seam instead of through claiming. Deferring work past a fan-out means owning which fan-out it belongs to. The chain's subscriber and the expander's each run exactly once per fan-out, so counting the first and matching it off in the second identifies the nesting level with no new seam in merged Stage 4a substrate.

  • Round 12 found round 11's counter in the wrong place. It counted invocations of the EXPANDER, which is optional: a lower-priority consumer can claim and stop the chain before the expander runs, while that fan-out's deferred subscriber still runs — so the nested pass went uncounted, looked outermost, expanded early, and outer pairing resumed with an invalidated record. The count now comes from a no-op consumer at the MINIMUM priority, which runs first in every chain invocation that reaches any consumer, and degrades safely: the only thing that can skip it is a claim ahead of it, which skips the expander too. A subscriber registered beside run_deferred cannot serve — the whole nested fan-out completes inside the outer chain's subscriber, before it would run.

  • Rounds 10–12 share a shape worth naming. Each fix was correct about the failure it was shown and wrong about the boundary of the mechanism it leaned on — first the chain's copy semantics, then its re-entrancy, then its short-circuit. A queue that outlives the thing that filled it has to name that thing, not approximate it.

  • Undo is cross-peer-degraded on CRDT frontends and that is ACCEPTED, named in the module header (Q#LN21): six source-peer optimistic inserts replaced by one daemon-peer op. set_round_trip_input would fix it and also disables dispatch_idle, so RET would stop inserting a newline.

Journey Stage 1a — PR #182 OPEN, review round 1 closed

  • Framing docs/journey-stage1a-framing.md rev 8 (four review rounds, two correction revisions found during implementation, one from review round 1 of PR #182). Branch journey-stage1a-directory-open, based on githubsucks/main @ 42025e4 (rebased onto 74301d1, then integrated 42025e4 and the landed-docs work below by merge — the branch is under review, so its history is no longer rewritten).
  • Recovery: git fetch githubsucks && git checkout journey-stage1a-directory-open. Everything below is committed and pushed; nothing depends on a worktree or /tmp.
  • Ships: the directory arm on resolve_target_buffer, EditorState::open rewritten as a caller of it (the unification), the path.open-directory chain + pmacs.path.directory_handler fallback slot, pmacs.window.commit_to with its scoped frontend and preflight, the nonconstructible destination userdata, the daemon bootstrap arm, and tests/journey_acceptance.rs (24 pins as of rev 8 — a count, not a constant; re-read it rather than quoting this line). No protocol change — still v20.
  • Doc updates ride the PR per COHERENCE §25: §2 grade + step-3 verdict row, §20 Priority 1 + the arc list, the GPU initial-target framing's Q#GT6 / acceptance 10 supersession, handoff §1.
  • Bite results (each mutation run against the full suite): scope stops swapping core.active_frontend → N6a + P3 fail, nothing else; preflight moved after the callback → P1 + P2 fail, nothing else; drop the ScopedFrontend arm from acting_frontend → N4b fails, nothing else. That last mutation is why N4b exists — it left N4 green. Round 1 of PR #182 added two more: dired's display back to switch_buffer, and prev read from the ambient window → each fails N4c alone. The scope pins the frontend, not the window — every routing pin before N4c varied frontend identity and none varied the selected window within one frontend, so 23 green pins missed it.
  • Ordering: PR #177 MERGED (2026-07-26), so 1a was unblocked. 1a lands before dired Stage 2. When 1a lands, Stage 2 must re-scout and revise its framing around the scoped pmacs.window.commit_to boundary before its implementation branch is cut. That revision is a prerequisite, not a review-time discovery.
  • Named deferrals carried out of this stage: dired's interactive paths (C-x d, tree descent, refresh) still rely on the ambient frontend a tick later and are not migrated onto captured destinations; the stale startup scratch buffer is still not removed (only the false doc comment is corrected); resolve_target_buffer's directory arm has no picker, only the chain that leaves room for one.

The CRDT half of the test corpus is dark in CI — NEEDS A LANE

  • No branch, no framing yet. Found while gating #166, then measured properly during the vterm as-framed audit. Deliberately kept out of #166 so a CI change would not arrive after review approval.

  • Root cause: .github/workflows/ci.yml never enables the crdt feature anywhere — zero hits across the workflow directory. The test job runs cargo test --all-targets --no-default-features --features luajit|lua54. Every #[cfg(feature = "crdt")] test is therefore not compiled in CI, not merely skipped.

  • Measured, --list under CI's exact flags versus the same flags plus crdt: 3,176 vs 3,449 — 273 tests dark. Re-measured at 74301d1 (2026-07-26; at fe8b8ba it read 3,170 vs 3,443, the same 273 dark — #176 added six tests, none of them crdt-gated). The number moves with every merge and must be re-measured, not quoted. #168 reported 3,024 vs 3,288 — 264 dark, 177 in the library — at 1b6a084; #178 then added CRDT-only generated-buffer coverage, and other lanes landed CRDT tests in between. Per target:

    dark CI full target
    185 1,848 2,033 the library itself (src/lib.rs)
    21 15 36 m5_5_acceptance
    13 1 14 gpu_invocation_acceptance
    13 1 14 gpu_initial_target_acceptance
    8 0 8 m10_11_acceptance
    6 0 6 auto_pair_crdt_acceptance
    6 0 6 m10_2_perf
    4 5 9 vterm_stage3_acceptance
    4 0 4 m10_10_perf
    3 0 3 compile_mode_crdt_acceptance
    2 22 24 theme_faces_acceptance
    2 0 2 m11_5_semantic_acceptance
    1 14 15 terminal_copy_mode_acceptance
    1 9 10 vterm_stage1_acceptance
    1 7 8 statusline_segments_acceptance
    1 10 11 gpu_font_acceptance
    1 0 1 auto_indent_crdt_acceptance
    1 0 1 m10_11_perf

    The rows sum to 273; the table is the whole census, not its head.

  • The single worst line is the library. cargo test --lib --features crdt is a REQUIRED local gate in CLAUDE.md, and CI has never run it. 185 library tests — the whole CRDT half — are developer-machine-only, and that count grows with every merged branch that adds a crdt-gated unit test.

  • Ten suites run zero or one test in CI, including gpu_initial_target (#148's entire acceptance, 1/14), gpu_invocation (#141's, 1/14), and a37, the Vterm Stage 3 real-daemon/real-PTY/real-wgpu path that #135 built specifically because "a decoded-message fixture would prove none of the three fit together".

  • ⚠ a37 will report green in the new job without running, unless the job builds pmacs-gpu AND sets PMACS_REQUIRE_GPU=1. Measured 2026-07-26 while gating #173. a37_real_daemon_real_pty_and_headless_gpu_ render_one_terminal_session derives its sibling binary path from CARGO_BIN_EXE_pmacs, and on a missing binary it eprintln!s a skip and returns ok. A fresh worktree running cargo test --features crdt --test vterm_stage3_acceptance reports 9/9 in 0.17 s having never run it; a real run takes ~4 s. Only PMACS_REQUIRE_GPU=1 promotes that skip to a failure, and CLAUDE.md applies that flag to cargo test -p pmacs-gpu — a different package, so the required local gate does not cover a37 either. The gpu-render job already sets the flag, which is what makes fix-shape part 2 sound; state it as a requirement of that job rather than inheriting it by luck, because a crdt leg added to the plain test job would run a37 vacuously.

  • a37 is also load-sensitive, which changes how to read the expected first-run failures. It passed at d152120 and failed at that same commit twenty minutes later, with a second agent saturating the machine with rustc in between; it then failed identically on d152120, 04c5ad1, and the #173 merge commit, which is how #173 established the failure was not its own. The signature is last_frame_text all spaces with rendered_nonuniform_frames nonzero — frames arrive, content does not. pmacs-gpu's own suite flaked the same way under the same load (201/202, then 202/202 on immediate rerun). So a red a37 on the first CI run is ambiguous by construction: before treating it as a real failure, run the same command on the merge base, and prefer serialized execution for this suite over retry-until-green.

  • Sort deliberate from accidental before proposing a fix. Some of the 264 are perf suites that are #[ignore]d by default and belong to their own jobs (m10_2_perf 6, m10_11_perf 1). m10_10_perf has no #[ignore] and no CI job naming it, so it looks accidental. This classification is not finished and is the lane's first task.

  • Fix shape, two parts (the flag combination is verified to work: --no-default-features --features luajit,crdt lists 10 vterm Stage 1 tests versus 9 without):

    1. a crdt leg on the test job for the non-GPU suites and the library;
    2. the GPU-requiring crdt suites onto the existing gpu-render job, which already has lavapipe and PMACS_REQUIRE_GPU=1 — vterm_stage3_acceptance, gpu_invocation_acceptance, gpu_initial_target_acceptance, gpu_font_acceptance.

  • Expect first-run failures, and budget for them. These would execute in CI for the first time ever: real PTY timing on CI runners, wgpu under lavapipe, and daemon-socket tests at unfamiliar concurrency. Start ubuntu-only and decide about macOS from evidence. A red first run is the lane working, not the lane failing.

  • Mitigating fact, verified rather than assumed: #166's three unit pins are not crdt-gated and do run under CI's exact flags, including the controller-release pin whose only job is catching the plausible wrong fix.

  • This lane also owns a --lib --features crdt flake, observed and scoped without overclaiming its cause (inherited from #178's gating, where the terminal lane recorded it). cargo test --lib --features crdt failed ~1 run in 5 on process::tests::setsid_escapee_is_not_reaped_and_teardown_reclaims_readers — active_reader_probe returning None at process.rs:3179 ("live runtime probe"). Pre-existing and unrelated to #178: that branch did not touch src/process.rs at all, and the test passed 10/10 standalone; the observed failures were during parallel full-suite runs. That localizes the trigger to suite load or interaction, but does not distinguish parallelism from another full-suite effect — no serial full-suite bite was run. The leading code-path explanation is the known drain_until trap: draining for Started also ticks, and a tick can reap the leader before the following active_reader_probe. That is an inference from the failure site and control flow, not yet a falsified root cause. Discriminating it belongs here. Two unnamed CRDT failures in #178's round-2 gating are a plausible match but remain unattributed — no test names were captured.

  • A second standing obstacle for this lane: cargo clippy --workspace --all-targets --features crdt -- -D warnings fails on main — measured at 74301d1: seven errors before the build aborts, four in src/daemon.rs (useless_conversion at 3996, missing doc backticks at 4076, too_many_lines 112/100 at 4083, an unneeded mut at 4965) and three in tests/vterm_stage3_acceptance.rs (too_many_lines at 637 and 793, a redundant continue at 843). Treat that as a lower bound, not an inventory: Clippy abandons the remaining targets once one fails, and a run on an older tree surfaced a further doc-backticks error in tests/auto_indent_crdt_acceptance.rs:42 that this run never reached. The standing gate list runs Clippy without crdt, so these lints have never been enforced. Any CI job that compiles the crdt targets has to fix them first or it will be red on arrival.

Bottom-panel lane (Arc 7) — Stages 1, 2A + framing MERGED; 2B is next

Stage 1, the Stage 2 framing, and Stage 2A are all on main. Stage 2B has not started.

  • Stage 2A MERGED as #177 (main @ 0a3fcd1, 2026-07-26, all twelve checks green at 8424172, three review rounds). Branch githubsucks/bottom-panel-stage2a and worktree ../pmacs-bp-stage2a are retained and carry nothing unmerged. Five commits: the classified census routing, the painter extraction + acceptance, the lane record, then the round-1, round-2 and round-3 review fixes. No protocol change; no behavior change for any frontend today — with panel_capable = false for semantic sessions, primary_document_window returns view.active in every existing configuration, so this is seam adoption that becomes load-bearing in 2B.

  • Stage 2B is approved and unstarted. It branches from main, not stacked on 2A, per the framing §9. Scope: protocol v21, the daemon panel projection, the GPU band, and the negotiated panel_capable flip. docs/bottom-panel-stage2-framing.md §7.2 carries its five acceptance criteria (A2B-1..5) plus the reassertion of parent criterion 52, and §8 records no open items, so 2B needs no further framing round. Its sharpest trap is §5.3's three-boundary split: the GPU text_area_bottom is status_band_top, geometry_capacity_bottom and document_text_bottom at once, and a blanket rewrite moves the status chrome along with the document while still satisfying an "everything moved" assertion — hence A2B-4's contrast form.

  • Verification on the merge result: cargo fmt --check clean; strict workspace Clippy clean; 1,832 default + 2,015 CRDT library tests; bottom_panel_stage2a_acceptance 17; bottom-panel Stage 1 46; statusline segments 8 CRDT; m11_5 semantic 2 CRDT; GPU initial target 14 CRDT; terminal config 12 CRDT; vterm Stage 1/2 10 / 6; folding Stage 2 48; M4 121; required GPU 202; git diff --check clean.

  • Every routed producer is now pinned at a seam its production caller uses, and each pin was falsified by revert: #1 follow, #2 lazy CRDT upgrade, #3 CursorByte, #5 decorations, #7 Viewport (aligns without focusing), #8 Pointer (aligns and focuses), #9 the terminal-context gate, #12 statusline, #21 the publication filter, plus the focus-class negatives. #1/#3/#21 required extracting three named helpers, because their only production caller is dispatcher_loop, which no test can drive.

  • Three lessons about the TESTS, not the code, all from review: (a) a structural test comparing the two authorities directly does not catch a misrouted consumer — only consumer-level assertions do; (b) a daemon-path test must register_session or the event is dropped at the uninstalled-session check before reaching the code under test; (c) a discriminating fixture must make the two routings DISAGREE — comparing two non-terminal buffers, or two windows with no selection, yields the same answer either way and proves nothing. Round 2 found four of my own pins vacuous by exactly these shapes, and round 3 found two more problems of the same family: a pin placed at a HELPER while production called it from a producer (reverting only the producer's call site left every test green), and a socket-pair assertion whose blocking read made a regression HANG instead of fail. Both now assert at the producer, with read timeouts on every read.

  • Review round 1 closed: 4 P1 + 2 P2, all real. The P1s were a stale-Pointer focus steal (the failed-alignment arm returned the window, so #8's activation focused it before dispatch_pointer rejected the buffer), the missing A2A-2 two-context fan-out, a census suite that asserted the AUTHORITY rather than the CONSUMERS, and the missing main integration. Two of the new pins were themselves vacuous on the first attempt — the dispatcher test passed because an unregistered session is dropped at daemon.rs:1962 before reaching the aligner, and the painter test was a fixed-point check that survived deleting text_view.render. Both now fail under their own bite.

  • vterm_stage3_acceptance::a37 is a pre-existing flake here, not a Stage 2A regression: measured 6/8 failures on the base commit and 7/8 on the branch in matched isolated samples. It needs a real daemon + real PTY + headless GPU and is documented load-sensitive. It also silently returns ok unless pmacs-gpu has been built, and is crdt-gated so CI never runs it at all.

  • Two suites are dark without --features crdt: m11_5_semantic_acceptance reports 0 tests and gpu_initial_target_acceptance reports 1 in the default config. Both are semantic-census suites, so Stage 2A must be gated with the feature on or its most relevant coverage never executes.

  • Stage 1 merged as #155 (main @ e745068, 2026-07-24, after two review rounds). No protocol change. Durable substrate facts live in docs/agent-handoff.md §1; the two round lessons are in §5.

  • Landed-docs follow-up merged as #156 (main @ d152120, 2026-07-25).

  • Stage 2 framing: docs/bottom-panel-stage2-framing.md revision 4, on branch githubsucks/bottom-panel-stage2-framing (three commits, one per revision), worktree ../pmacs-bp-stage2, based on githubsucks/main @ ccf29e3. Round 1 closed 2 blocking + 3 high; round 2 closed 1 blocking + 2 high + 1 medium and decided both open items; round 3 closed 1 blocking + 1 high + 1 medium. No open items remain. The approved parent framing docs/bottom-panel-framing.md (rev 4) remains authoritative, including its acceptance criteria 37–55.

  • Retained, carrying nothing unmerged: branch bottom-panel and worktree ../pmacs-bottom-panel.

  • Stage 2 ships as two serial slices, 2A landing before 2B branches: 2A = classified §1.3 census routing + paint_frame per-window painter extraction (with the active-window auto-scroll preparation), no protocol change; 2B = protocol v21 (InstanceMessage::PanelFrame plus FrontendEvent::{FrontendCellGeometry, PanelResizeRows, PanelPointer}, gated both directions, each extended enum byte-pinned on its own previous final variant), daemon panel projection, the GPU band, and the negotiated panel_capable flip. Stage 3 is the adopter default flip.

  • Correction — this entry previously mis-stated the census contract. It is not "route every consumer through primary_document_window". Q#BP14 classifies the 23 reads into four classes and routes only the Projection class that way; focus/input (#13–#15, #23), focus chrome and surface-routed (#16–#19), and focus/session (#20) keep their own authorities. Rerouting them would break remote-op validation and application, DispatchIdle, presence, focused search/menu/completion routing, and terminal bell ownership. The Stage 2 framing carries the full table.

  • The GPU document bottom is three boundaries, not one. text_area_bottom (pmacs-gpu/src/main.rs:8490) is today status_band_top, geometry_capacity_bottom, and document_text_bottom at once. Once a band is installed they diverge: the status chrome must stay pixel-identical at the physical window bottom while document consumers move. A blanket rewrite of that helper moves both together and passes an "everything moved" assertion, so the Stage 2 criterion asserts both directions in one scenario. The census is 20 production sites (8 status-owned, 12 document-owned) + 1 definition + 8 test sites = 29 matches; the framing carries the per-site table. The three easiest to misclassify are document completion :6140, minibuffer candidates :7351, and edge scrolling :8561 — each with its own visible symptom.

  • Folding Stage 3 and this arc's Stage 2 both touch the semantic projection. Whichever is framed second re-scouts the other's landed state.

Folding lane (Arc 6) — Stages 1 and 2 MERGED; Stage 3 (GPU) is next

Both shipped stages are on main; nothing in this arc is in flight. Stage 3 has no branch and no framing yet.

  • Stage 1 (headless fold engine) merged as #142, Stage 2 (grid/daemon collapse) as #149 — both under "Closed since the last snapshot".
  • Retained, carrying nothing unmerged: branches folding / folding-tui and worktrees ../pmacs-folding / ../pmacs-folding-tui. The framings docs/folding-framing.md (rev 5) and docs/folding-stage2-framing.md (rev 4) are the approved artifacts Stage 3 re-scouts against.
  • Stage 3 (GPU) obligations, already named by the framings — the starting point for its own framing doc: GPU collapse at TUI parity; caret/hit-test fold-awareness; the BufferSnapshot fold-mirror clear (parent R2-4 — without it, empty-after-revert diff suppression leaves stale folds on the GPU, the same trap class as #120); CRDT-origin and GPU-optimistic interactive unfold (parent R2-3); and flipping FrontendView.fold_projection to true for semantic frontends, which Stage 2 deliberately left false (Q#FD21).

Parked lane: kill-ring browser + persistence

  • Portable branch: githubsucks/kill-ring-browser
  • Parked framing head: 503c489
  • State: framing only, revision 2; no implementation and no PR.
  • Status: explicitly parked by the user on 2026-07-20.
  • Its original scout was based on 0efb5cd. The preserved framing marks this ground truth stale and requires a complete re-scout against the then-current githubsucks/main before implementation.
  • Compile-mode has merged since the original scout, so old “compile-mode in flight” keybinding/touch-set assumptions are not authoritative.

Recovery worktree, only when the user un-parks it:

git worktree add --track \
  -b kill-ring-browser \
  ../pmacs-kill-ring-browser \
  githubsucks/kill-ring-browser

Documentation lane

  • Portable branch: githubsucks/handoff-2026-07-20
  • Carries synchronized AGENTS.md / CLAUDE.md, this ledger, the durable handoff refresh, and the keybinding reference correction.
  • It changes no runtime code.
  • Review and merge this documentation branch separately; it must not be folded into a feature framing branch.
  • Now also absorbs both landed arcs: Vterm Stage 1 (#126) and the config registry (#127). Canonical main is merged into it up to 2e37c04, so its diff against main is documentation only.

Closed since the last snapshot

  • Terminal configuration + copy mode arc — BOTH STAGES MERGED, lane removed. Stage 1 #173 (main @ cf54270, one review round) and Stage 2 #178 (main @ fe8b8ba, four review rounds, twelve checks green on head 1b44c69 — verified by head_sha, not by the check summary), both 2026-07-26, both with no protocol change. Approved framing: docs/terminal-config-and-copy-mode-framing.md rev 4, committed as the first commit of Stage 1's branch; its Q#TC6a carries a superseded-in-part box rather than a silent rewrite. Durable facts moved to docs/agent-handoff.md §1 (the arc bullet) and §4 (the set_generated_contents invariant) per rule 3 below, and to COHERENCE.md §14. Stage 2 ships eight of nine criteria and the missing one is named — criterion 17 needs a real GPU frontend, so it waits on the a37 footing; the handoff records what it must assert. Branches githubsucks/terminal-config and githubsucks/terminal-copy-mode with worktrees ../pmacs-terminal-config and ../pmacs-terminal-copy-mode are retained. The gate-run flake found while gating #178 moved to the CI crdt-coverage lane above, which owns its discrimination.

  • Dired Stage 1 (the directory view) — MERGED as #165 (main @ c8ec8f3, 2026-07-25, after one review round). pmacs has a directory surface: C-x d / C-x C-j, one read-only buffer per directory named *dired:<canonical path>*, a dired major mode carrying RET/f, ^, n/p, g, q, s. No wire change (v20). The Rust is two things — a per-entry-tolerant read_dir (Q#DR6), which had to be Rust because read_dir_blocking fails a whole listing on any of five per-entry conditions and a tolerant wrapper cannot be written in Lua at all, and normalize_buffer_path going pub as pmacs.path.canonicalize (Q#DR2's preferred end state, so no Lua mirror exists and Stage 2 owes no mirror removal). The frozen m8_1/m8_2/m8_3 counts are unchanged, which is the additivity gate. 15 claims bite-verified; one came back VACUOUS (acceptance 3c cannot pin descent routing — dired holds focus in its own panel, so dedication is the only discriminator) and is documented at the assertion rather than relabelled. Its branch (dired-stage1) and worktree (../pmacs-dired-stage1) are done; the abandoned dired branch (ffdd642, ../pmacs-dired-arc) was superseded by a fresh cut and carries nothing unmerged. Stage 2 (marks and operations) and Stage 3 (wdired) each still need their own framing, and the frozen fixture shrinks after Stage 3. Durable substrate facts and both new ops lessons live in docs/agent-handoff.md §§1/5; the implementation notes are docs/dired-framing.md §0, S1-1…S1-12. Two named forward items for Stage 2: apply_resource_op's rename rebind is exact-PathBuf-equality, first-match-only, looked up with the raw path while stored paths are normalized — so a directory rename strands every buffer under it, and pmacs.fs.rename has zero production callers, so it can be fixed at the primitive; and Q#DR5's seam is the main-thread drain AsyncRuntime::tick, not _take_result, where rename settles as an undifferentiated ReplyKind::FsUnit and so must be keyed on JobKind::FsRename.

  • GPU terminal input (the double terminal-layout sync) — MERGED as #166 (main @ b889873, 2026-07-25, one review round, all twelve checks green after a macOS PTY-timing rerun). The dispatcher applied both 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. sync_terminal_layout is now split into a frontend-kind-neutral half (panel reconcile + controller liveness) and a grid-only geometry half, with the loop body extracted to sync_terminal_layouts_for_tick so the exclusivity is structural. No protocol change (v20). Durable lessons are in docs/agent-handoff.md §5; the framing (docs/gpu-terminal-input-framing.md rev 2) carries three falsified hypotheses, the two-pre-image bite matrix, and two named out-of-scope items (Q#GT5 interactive-shell echo on a raw PTY, which reproduces in-process and so is not the GUI/TUI asymmetry; and a geometry change appearing to clear the visible screen, which reproduces pre-fix). Branch gpu-terminal-input and worktree ../pmacs-gui-term-input retained. Its landed-doc pair MERGED as #168 (main @ 1b6a084, 2026-07-26): #166 recorded as landed, the CI crdt-coverage gap measured (264 tests dark workspace-wide, 177 in the library — a reading taken at 1b6a084 and kept here only as history. The CI crdt-coverage lane above is the authority for the live figure; do not quote this one forward), the vterm audit corrected — "only 3 of 9 acceptances drive a real daemon" was optimistic; without the frontend binary the honest number is 2 — and the a37 findings folded into the coverage lane.

  • Inline-math slice — MERGED as #158 (main @ 5aa9044, 2026-07-25). Detect → parse → layout → draw for $…$, entirely inside pmacs-gpu, no protocol change. Verified by the user's manual pass on a real paper after the landing. What is worth carrying forward:

    • The v0 subset is 34 Greek symbols, sub/superscript, and \frac. An unsupported command fails the whole span back to source, so on a real document most inline spans still show LaTeX. Widening the symbol map is the highest-value next increment — ahead of display math, which is also deferred.
    • A stale frontend binary is invisible from the source tree. The slice lives only in pmacs-gpu, so after it merged the feature was absent until cargo build --release -p pmacs-gpu and a client restart; the daemon needs neither. Diagnose with strings on the binary (Latin Modern Math, MathBox) rather than by re-reading the checkout, which was already current.
    • Main was integrated three times in one day (8c86d34, 46a1b8f, b889873), merged not rebased to preserve review anchors. Two conflicts, both this ledger and nothing else. The dangerous case was the one that did NOT conflict: #166 auto-merged into pmacs-gpu/src/main.rs, the file this lane rewrites, because the two edits sat in different regions of it. Decide whether to integrate from the shared-file set, never from whether git complained.
    • Integration was proved by test-count reconciliation, not by a green run: predict what the other side adds, then check the deltas. GPU 199→202 matched e547a90's 3; later lib 1,826→1,829 and CRDT 2,003→2,006 matched #166's 3, with GPU unchanged because #166 adds none. Suites 91→92 was #161's new binary.
    • Why the branch had no CI for a day: a conflicting PR builds no merge ref, so no pull_request run is created. The ledger previously recorded this cause as unidentified; it is not. Check mergeable and confirm a run exists for the current head SHA.
    • m4_5_basedpyright has no timeout and hangs forever, parking a --workspace sweep (observed 2h26m at 38 of 92 suites). It is intermittent, so an earlier clean sweep proves nothing. Sweep with cargo test --workspace --no-fail-fast -- --skip basedpyright and judge progress by whether the suite count advances.
    • Named v0 approximations: the peer-caret half of acceptance 14 is pinned at the mapping level, not pixels; a soft-wrapped spacer draws its box whole at the first run's origin; the fit budget reads the bundled code face even under a custom set_font family.

  • Bottom panel Stage 1 — MERGED as #155 (main @ e745068, 2026-07-24, after two review rounds). Window placement, window parameters, TUI side windows, the divider, and the adopter display opt-in, with no protocol change. Both rounds found the same class of defect and are worth keeping:

    • Round 1: the Q#BP6 side-window split guard had no production caller — C-x 2 still reached plain split_active — and survived because the acceptance test called the core method directly.
    • Round 2: Q#BP7's terminal growth re-arm had never been implemented, and the assertion meant to pin it (at_bottom) is a geometric readout that a still-anchored view satisfies; the anchor assertions beside it compared "" with "" because the PTY fixture emitted LF-only output.
    • Post-round-2 self-review, caught by CI going red on all four Test jobs: resolving pmacs.window.buffer()'s no-argument arm through the acting frontend made a total function partial, and six runtime modules silently dropped operations (kill_ring_acceptance 30/30 → 25/5). Fixed in 9110f9f before merge.
    • Gating fact found on the way: the workspace sweep must run with an isolated XDG_CONFIG_HOME, because the real user init.lua installs a local package and the losing race leaks a status message into painted-frame comparisons. There is also a latent pre-existing main bug in the buffer CRDT undo path, unrelated to this arc.
    • compile_mode_acceptance is load-sensitive under default parallelism (~1 run in 3, a different test each time); verified pre-existing by swapping in main's compile.lua. It is 67/67 at --test-threads=1.

  • GPU initial target — MERGED as #148 (main @ 0dd16a5, 2026-07-24, after two review rounds). pmacs --gpu [--socket …] FILE opens a target before the GPU window appears. Protocol bumped 19 → 20: a semantic-session SessionBootstrapRequest after AttachRequest, plus an appended InstanceMessage::InitialTargetResult pre-window readiness barrier; v6–v19 wire encodings are unchanged. Root owns launcher tilde/cwd resolution and exact raw-byte path transport; the daemon resolves/dedups/loads the target and runs load/switch hooks inside one dispatcher transaction, then publishes CRDT-upgraded targets to existing grid replicas (gated on upgraded_to_crdt, independent of the load/create outcome, so a dedup onto a hidden not-yet-backed buffer still reaches pre-attached replicas — round 2 finding). Semantic replicas receive a publication only when displaying that buffer, so a second target launch cannot switch an existing GPU window. Round 2 also closed a failure-containment gap: every dispatcher-side bootstrap failure now shuts down the socket (a dropped write-half clone does not close a shared FD), and the dispatcher drops any event from a session that was never installed, rather than reaching absent render/size state. Integrated cleanly with Folding Stage 2 (#149): fold projection at attach is selected from the same negotiated semantic_render bit the target bootstrap uses. Its lane, worktree (../pmacs-gpu-initial-target), and branch (gpu-initial-target) are done; the -framing branch is kept. Durable substrate facts and both review-round lessons live in docs/agent-handoff.md §§1/5 and docs/gpu-initial-target-framing.md rev 3.

  • Folding Stage 2 (grid/daemon collapse) — MERGED as #149 (main @ 6ed4fe9, 2026-07-24, after five review rounds). The grid TUI now renders collapses. Spine (Q#FD12): src/fold_view.rs's VisibleLineMap, derived from the fold store plus a window's line offsets and never stored, threaded as Option<&'a VisibleLineMap> on a lifetime-bearing Viewport<'a> that stays Copy. No wire schema or protocol change; the GPU path is Stage 3. 48 acceptance tests on the real paint_frame grid, every behavioral claim bite-verified. Durable design points, each a trap Stage 3 inherits:

    • 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 a 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, since a wheel event names a pane without activating it;
    • fold projection is per-frontend (FrontendView.fold_projection) — 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.

    Process notes worth keeping: main moved under the arc, and the merge was textually clean but not semantically clean (#146 added Viewport literals the new folds field invalidated) — a clean git merge-tree does not mean the merged tree compiles. CI was red at review on the macOS/luajit outline_5_level_100_entry_renders_within_100ms budget flake and went green on rerun.

  • Documentation ledger refresh — MERGED as #147 (main @ 0a479ae, 2026-07-24). The #142 housekeeping, expanded after review found the ledger stale through four merges rather than one. Its own macOS/luajit red was the vterm VTERM_ALT_READY PTY timeout; green on rerun.

  • Web grammars HTML + CSS — MERGED as #146 (main @ 47581f4, 2026-07-23). .html/.htm/.xhtml and .css highlight off the official tree-sitter-html 0.23 / tree-sitter-css 0.25 crate query constants (no in-repo overlay), and HTML's INJECTIONS_QUERY lights up <script> → js and <style> → css. Durable lesson recorded in docs/web-grammars-html-css-framing.md: the highlight.rs capture table is global, so adding a capture name retro-paints every other language — check the reverse direction and pin it.

  • LaTeX Stage 1 — MERGED as #144, with its parent inline-math framing committed as #145 (main @ f09b0a1, 2026-07-23). .tex/.latex/.sty/.cls highlight via codebook-tree-sitter-latex 0.6 plus the first in-repo query overlay (builtin/queries/latex/highlights.scm, include_str!) — the reusable pattern for grammars whose crate ships no usable queries. The crates.io tree-sitter-latex is provably broken (no scanner.c). The math parser and Tiers 3–4 are deferred to the inline-math arc.

  • Folding Stage 1 (headless fold engine) — MERGED as #142 (main @ c49a8c7, 2026-07-23, after three review rounds; round 3 clean). The instance-side fold store + translating/dropping View, the structural source (derived head line, closer-aware tail), the Lua data API + interactive C-c @ commands, the command-path pre-edit unfold, and authoritative-empty FoldState production landed with no protocol bump. The folding branch and worktree (../pmacs-folding) are retained but carry nothing unmerged; the folding-framing.md framing is preserved. CI red at merge was an unrelated environmental perf flake (outline_5_level_100_entry_renders_within_100ms, macOS/luajit only), green on rerun. Stage 2 has since merged as #149 (above); durable substrate seams live in docs/agent-handoff.md §1.

  • Vterm Stage 3 (protocol v19 + GPU terminal) — MERGED as #135 (main @ cac4961, 2026-07-22, after two review rounds). Arc 5's terminal stage is complete (compile mode #113, Stage 1 #126, Stage 2 #130, Stage 3 #135). Its lane, worktree (../pmacs-vterm-gpu), and branch are done; durable substrate facts live in docs/agent-handoff.md and docs/vterm-framing.md.

  • Branches deleted 2026-07-22 (authorized): vterm-stage3-framing (Revision 8 framing; its content is carried on vterm-gpu, verified as a superset before deletion — the branch was NOT an ancestor of vterm-gpu because the framing was copied rather than merged, so it needed a forced local delete) and tab-width-parity (a clean ancestor of main via #137). Both removed as worktree + local ref + githubsucks ref; the origin tracking refs were pruned. The -framing branches for each are deliberately kept.

  • Tab-width rendering parity — MERGED as #137 (main @ 2625ec7, 2026-07-22). One fixed 8-column TAB_STOP_COLUMNS in pmacs-protocol now drives core/TUI columns, GPU code projection, and minimap width; source bytes and protocol ranges are unchanged. Its lane, worktree, and tab-width-parity branch (local + githubsucks) are deleted; the tab-width-parity-framing branch is kept. This closes the long-standing "tab width is a rendering-parity bug, NOT a config gap" deferral recorded in docs/agent-handoff.md §5.

  • Locals-query processing — MERGED as #134 (with handoff #136), and modeline detection handoff #133. Both landed between this lane's base and its canonical-main integration.

  • Config registry — MERGED as #127 (main @ 2e37c04). Its lane (config-registry, worktree ../pmacs-config-registry) is done; the branch is kept but carries nothing unmerged. Durable substrate facts moved to docs/agent-handoff.md §1 per rule 3 below.

  • Both this and Vterm Stage 1 ran as concurrent lanes in sibling worktrees off main, with the shared files (src/editor.rs, src/lua_bindings/mod.rs, src/lib.rs) assigned to one lane each in advance. The rebase of the second lane onto the first had zero conflicts — worth repeating for future parallel work, along with its precondition: agree the file split before either lane starts, and keep each lane's footprint in the other's files to a single line.

Update protocol

Whenever a listed lane changes materially:

  1. update its public branch and head/state here;
  2. record new verification and remove superseded caveats;
  3. keep durable architecture in docs/agent-handoff.md, not here;
  4. remove the lane after merge or abandonment;
  5. verify every recovery command from a clean worktree before calling the transfer complete.