The 4a/4b split held; five P1s against rev 6's own content, all real,
all reproduced. Four share a root: rev 6 verified its external facts and
under-verified its internal ones.
1. Stage 4a's declared footprint excluded the tests its own acceptance
required. 46a-46e cannot live in tests/auto_pair_acceptance.rs,
which criterion 46 requires byte-identical. Footprint now names
tests/typed_edit_chain_acceptance.rs and gates on it.
2. Pending abbreviation state had the wrong owner. pmacs is
multi-frontend: EditorCore.views is per-FrontendId with its own
active window, take_typed_edit is already frontend-keyed, and
buffer.after-switch fires with no arguments — so a buffer-keyed
clear-on-switch lets any frontend discard another's pending
abbreviation. Now keyed (frontend, buffer) with a window check,
frontend-scoped clearing, a frontend.detached purge, and acceptance
45i, which the buffer-keyed design passes every other criterion
without.
3. The shortest-match rule was missing its tie-break: upstream keeps
declaration order among equal-length shortest keys, and 101 prefixes
have equal-shortest candidates resolving to different symbols (f
picks f< over f>). A pairs-iterated Lua map cannot express this, so
the vendored artifact is now an ordered sequence and resolution sorts
by (#key, source rank). Rev 6 missed this because it declared the
package ships no README after a 404 on the package root, with the
directory listing showing src/README.md already in hand — a 404 on a
guessed path is not evidence of absence, and the README states the
rule in one sentence.
4. The generator's rejection rule rejected the current table: \ is a key
and " begins eleven, while acceptance 45d requires \ to work.
Replaced with canonical lossless escaping; aborts only on duplicate
keys, invalid UTF-8, and a failed self-round-trip. 45g no longer
claims to diff against abbreviations.json, which is not shipped.
5. Durable and volatile state were not reconciled. agent-handoff.md
anchored main at d152120 with neither #167 nor #170 and no Lean arc
bullet at all; active-work.md kept 407 lines of merged Stage 1/2/3a/3b
history against its own instruction to prune merged entries, under a
stale snapshot date. Durable facts moved to the handoff; the ledger
keeps only the unlanded Stage 4 lane.
Also corrected: 119 multi-codepoint symbols (26 with $CURSOR), not 93;
three backslash values, not two; Q#LN22 now states the terminating-\
reprocess rule acceptance 45d depended on; acceptance 38 says the
terminator is retained, so undo restores "\alpha " with its space;
coherence cites golden-journey step 5, not step 4; and the
config-registry prior art points at Q#LN22.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011LFvC4FQtux4y32KuevZ7B
Stages 3a and 3b landed (#167, #170). Re-scouting Stage 4 against main
@ d400f30 produced six findings that change the plan and three that
confirm it. The pmacs-side facts were verified in a worktree at that
commit; the upstream facts by reading leanprover/vscode-lean4 @ 17d1d08.
The split: Stage 4's risk column read "refactors pair.lua's provenance
read" — every language's auto-pairing — for a stage the prose called the
Lean input method, which is exactly the rule §4 states and exactly what
round 4 found for Stage 3. Rev 5 had noticed the shape and answered it
with a commit boundary; a commit boundary is not a review boundary.
Stage 4a is now the typed-edit consumer chain (substrate, no Lean) and
4b the input method.
Rev 5's expansion semantics were wrong in three ways. Resolution is the
shortest key having the input as a prefix (\al yields ∀ from `all`, not
`alpha`); there is no terminator list at all ('+ ' is a key, so space
extends after \+; '\' is a key, so \\ yields \); and an unmatchable tail
is appended rather than dropped (\alp7 yields α7).
Three further findings. There is no cursor-motion hook, so acceptance 43
as written was not buildable and abandonment is lazy. dispatch_key is
only half of 4b's production path — \ and the letters are not excluded
from the optimistic classifier, and that producer is crdt-gated, so a
crdt-gated integration test is dark in CI and dark in the gate list. And
the whole expansion has cross-peer-degraded undo, a wider bite than
Q#LN6's three bracket pairs; set_round_trip_input would fix it and is
rejected with reasons.
New decisions Q#LN21 (undo degradation) and Q#LN22 (the state machine);
Q#LN10 and Q#LN11 rewritten; §2.11 records the upstream algorithm; §9.1
states the coherence impact for both stages. Acceptance keeps its
existing numbers and adds letter suffixes on both sides of the split.
Citation sweep per COHERENCE §25: five live citations moved in the 50
commits since rev 5.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011LFvC4FQtux4y32KuevZ7B
Make command-time attachment healing cancel an armed terminal restart
before replacing the server, while keeping request-only lookup pure and
restart-safe.
Track config-driven server ownership privately, bound every fallback
server per SID, scope no-swap retirement to the failed root, and route
the shipped Lean diagnostics command through the safe resolver while
waiting for initialization.
Add direct acceptance counterexamples for all five review findings and
record the sixth-round verification and vacuity lesson.
Five findings, all real. The blocker and both majors are the same
mistake in three places: a claim asserted somewhere cheaper than where
it actually lives.
COHERENCE.md was stale in four places, not the three reported. Step 8
still read "no keybinding" and §11 still read "five settings", but §6's
dispatch table also still cited `is_terminal_escape_chord` — a symbol
this branch deletes. §25 requires that update to ride the PR, so a PR
changing audited ground truth has to re-grep the audit for its own
symbols, not only for its topic.
Acceptance 5 asserted a registry round-trip, which is a test of the
registry: it stayed green with the setting's only consumer deleted. It
now opens a real terminal whose child overflows the 24-row screen,
scrolls the view to its oldest retained row, and asserts LINE001 is
present at 10,000 and absent at 0.
Acceptance 8a waited for the session count to fall, which the rejected
editor-side cache map satisfies exactly — a map with no purge hook
leaks while sessions drain. Adds `TerminalManager::escape_caches()`, the
lifetime half of Q#TC4c's contract that `escape_parses` cannot cover.
`table.sort` over `pmacs.terminal.profiles` raised "attempt to compare
number with string" on the unknown-profile path whenever the user's
table held both a string and a numeric key, replacing the exact
diagnostic being asked for; `%q` raised likewise on a non-string
`profile` argument. Both are partial functions applied to user input on
a diagnostic path.
Also corrects the framing's status line, and a status message whose
embedded whitespace run had survived a rustfmt reflow.
Three new bites, each falsified by revert: deleting the scrollback
consumer fails acc5 and only acc5; restoring the raw-key sort
reproduces the comparison error verbatim; and implementing the rejected
map fails the new acc8a at left: 2, right: 1 while passing the old
session-count version.
Merges githubsucks/main @ ccf29e3.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016gGQC6eqHJVbZJ5Hg7aLer
Round 5 review: one P1, a frontend scope hole, and three P2s.
**1. A fallback that SPAWNS and then dies retried forever.** The
once-per-buffer guard bounds calls to `_attach_buffer`, not the server
those calls produce. `ensure_server` still never forwards `cfg.restart`,
so the fallback inherits `OnCrash`; an executable that exits before
`initialize` is respawned by the manager with no attempt ceiling —
silently, because `latched` has already disabled the primary's failure
poll. The fallback now gets its own one-shot die-before-initialize
watch, which retires it (ending the respawn loop) and reports.
The prior failing-fallback test used a NONEXISTENT executable, so it
only ever exercised synchronous ENOENT. To reach "spawned, then died"
the fixture has to actually spawn.
**2. Simultaneous frontends.** Both repair triggers read the ambient
`pmacs.window.buffer()`, and the daemon restores `active_frontend` to
the last-dispatched frontend before `tick_processes` — so a Lean buffer
active in ANOTHER frontend receives no `buffer.after-switch` here and
stays stale after its server is globally retired.
Fixed at the seam that is frontend-agnostic: **make consumption safe.**
`attached_for_active` now rebuilds rather than returning a record whose
server is dead, and `attachment_for_request` reports none (it must not
perturb LSP state, so it cannot rebuild). Whichever frontend runs a
command is the active one while it runs, so healing at the point of use
reaches every buffer no eager sweep can. This also closes the half where
a dead attachment was handed to a command and the request vanished.
**3. The retirement sweep stopped user-managed servers.** Selecting on
`language_id == "lean4"` also names servers the user spawned from
`init.lua`, which are not derived from `pmacs.lsp.config.lean4`. It now
keys on the `default-lean4` label `ensure_server` stamps — the
derivation discriminator.
**4. Repair ran even when no swap occurred.** `swap_to_fallback()`
returning false left `latched` true, so the next tick retried the
UNCHANGED configuration and reported it as a fallback failure. Split
into `probe.fallback_installed`: repair exists to apply a swap, so no
swap means nothing to apply.
**5. The once-per-buffer assertion counted table keys**, which cannot
distinguish "once per buffer" from "every tick for one buffer" —
cardinality stays 1 either way. Replaced with a numeric attempt counter;
the bite reports 174 attempts against the expected 1.
Five bites, each against 7c37bdc: no fallback watch -> attempt reaches
4; retire by language_id -> the user's server is stopped; gate repair on
`latched` -> a repair is attempted with no swap; drop the
once-per-buffer guard -> 174 vs 1; hand back a dead attachment -> a
command receives a `stopped` server.
Two more vacuity shapes recorded in the ledger (8 and 9): counting
distinct keys cannot bound repeated work, and a nonexistent executable
cannot reach any post-spawn failure.
`docs/agent-handoff.md` §1 still said Stage 2 "needs its own
re-framing". It is framed, so that line would be false on `main` the
moment this branch merges.
It now records the approved shape — protocol v21, two serial slices
(2A census routing + painter extraction, then 2B wire/projection/band/
capability flip), parent acceptance 37-55 still authoritative — and
carries the census classification rule itself, since that is the fact
the ledger previously got wrong and the one a future reader is most
likely to re-derive incorrectly.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Round 4 review: one P1, and it is the same defect for the FOURTH time.
`pmacs.lsp.config.lean4` is a single global entry, so swapping its
command invalidates **every** Lean buffer and **every** Lean server —
Q#LN15 gives one server per project root, so there can be several.
Rounds 1-3 each repaired one buffer and retired one server, and round 3
shipped "repair the armed target, strand the rest": status and config
said fallback while a second open Lean buffer stayed on the retired
command, and a second project root's server stayed live.
The shape that actually holds:
* **Retire ALL `lean4` servers on latch**, not the one the probe
happened to name. `probe.primary` identifies the server the VERDICT
is about; it was never the set of servers the swap invalidates.
* **Repair each buffer lazily and at most once**, when it becomes
active — on `buffer.after-switch` and on the tick. `_attach_buffer`
is an active-buffer-only seam, so a global swap cannot be applied to
every open buffer at once; it has to be applied as they surface.
lsp.lua's own `after-switch` re-pushes views but does not rebuild a
stale attachment, so nothing else covered this.
* The **once-per-buffer bound** is load-bearing: without it a fallback
that also fails to spawn would retry every tick forever — the
round-2 defect, which a naive global repair loop would reintroduce
for every buffer instead of just one.
* `shutting-down` is deliberately not treated as stale. It is still
live by `server_is_live`'s reckoning, so attaching would early-return
the stale record and burn that buffer's single attempt on a no-op.
P2: argument-inclusive attribution was implemented in round 3 but pinned
only by "contains the command name", so a mutation dropping every
argument passed. Now asserted against the exact `<command> <args>`
string.
Also fixed a vacuous assertion this refactor created: a test checked
`_probe.reattach_from == nil` for a field that no longer exists, which
reads as nil and passes for nothing. It now asserts a positive count of
recorded repair attempts.
Three bites, each against 73587b0: repair only the armed buffer -> the
second buffer stays on `lake`; retire only the named server -> one live
stale server remains; drop arguments from attribution -> the exact-string
assertion fails.
The ledger records a second durable lesson beside the vacuity one: **a
scope error repeats until the scope is named.** Four rounds of locally
correct fixes, none of which asked what the config swap invalidates.
When a change edits shared state, enumerate everything derived from it
before repairing anything.
Closes review round 3 — 1 blocking, 1 high, 1 medium.
**R3-1 (blocker) — the call-site table contradicted the source.** The
three-boundary model was right; five rows of its classification were
not, and each was a real defect:
- `:6140` is `completion_dropdown_layout` — DOCUMENT completion
placement, deriving the space below the anchor line. Classified
status-owned, it would let completion overlap the panel.
- `:7195` and `:7212` are the `status_buffer` / `status_left_buffer`
`TextBounds.top` — status text bounds, classified document-owned.
- `:7351` clips global minibuffer CANDIDATE glyphs to the dropdown's
band anchor; classified document-owned, they would be clipped
against a boundary the dropdown does not sit above.
- `:8561` (`edge_scroll_direction`, document edge scrolling) was
missing entirely, leaving it tied to the old bottom.
- `:8077` is `code_caret_rect_in_clip` — caret clipping, not
completion placement. Its class was right, its label wrong.
Every production site is now individually verified against the source
and tabulated with what it actually is. The census is stated as
arithmetic a reader can check: 29 matches = 20 production + 1
definition + 8 test sites.
Root cause recorded in the revision history: rev 3's table was built
from a `grep | head -20` over 29 matches, which is precisely why
`:8561` vanished. The minibuffer's status-owned status is now argued
from Q#BP14b rather than assumed — it is global, bufferless chrome
anchored to the status band, so all four of its sites stay with the
band.
**R3-2 (high) — clamps preserved.** The three equations permitted
negative coordinates on a surface shorter than its chrome, where
today's `text_area_bottom` clamps with `.max(0.0)`. All three now
clamp at zero, which keeps the "exact formula" exact exactly where it
matters most.
**R3-3 (medium) — attachment rejection classified SHARED.**
`validate_cells` also rejects `cell.attachment.is_some()`
(`terminal.rs:305`), whose error text reads "which terminals never
use" (`:190-191`) — phrased as a terminal-specific fact, which is why
rev 3's "exact split" missed it. Panels implement no attachment
rendering in Stage 2, so a `PanelFrame` carrying one describes a
surface the GPU would silently not draw; shared rejection fails closed
on the producer side instead. The message is reworded grid-neutral
when it moves, and giving panels attachment rendering later moves the
rejection back deliberately rather than by default.
A2B-4 now names the counts on both sides (twelve document-owned move,
eight status-owned do not) and carries the three symptom-bearing rows
that a plausible misclassification produces. §9 records that the GPU
three-boundary split belongs to 2B, not 2A — it is only observable
once a band can be installed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Round 3 review: two P1 asynchronous-correlation defects, with the
focused suite at 25/25 while both were live.
**1. A late version verdict retired nothing and claimed success.**
`probe.watching` is cleared the moment the server initializes — it is
failure-polling state. A slow `lake --version` landing after a
successful initialize therefore reached `fire_latch(nil)`, which retires
nothing: `_attach_buffer` found the still-live primary attachment,
early-returned it, and the retry counted that as done. Status said
"falling back", the config named the fallback, and the buffer stayed on
the old server.
**That is the round-1 silent no-op arriving through a third event
ordering** — first as "no re-attach at all", then as "re-attach cleared
by an unrelated buffer", now as "re-attach satisfied by the server we
were supposed to replace". The fix separates the two facts that were
being carried by one field: `probe.primary` is the server the verdict
applies to and survives initialization; `probe.watching` is the
failure poll and is cleared by it.
The existing fixture could not reach this ordering at all — its `serve`
sleeps, so the primary can never initialize before `--version` returns.
The new one execs the fake LSP for `serve` and delays 0.6s before
reporting 3.0.0.
**2. `buf_key` was the most recently loaded Lean buffer.** Written on
every Lean `buffer.after-load`, so a second Lean file opened before the
verdict became the rebuild target while the latch still watched the
FIRST buffer's server. Target buffer and primary server are one fact and
are now armed together, exactly once. Both files in the new test share a
package, so mis-targeting shows up as a stranded buffer rather than as
two unrelated servers.
**3. The failure message hardcoded `lake serve`** after the latch became
command-agnostic, telling a user whose `my-lean-wrapper` failed to go
debug lake. `configured_command()` names what is actually configured,
arguments included.
**4. The ledger** now records all fifteen bites across the three rounds,
both prior review rounds' findings (the round-2 block was lost when an
earlier edit script aborted before writing), and the durable lesson.
That lesson, recorded for the handoff: **six tests across three rounds
were written, ran green, and pinned nothing** — caught only by biting.
The shapes are enumerated in the ledger; the rule is that a test is not
evidence until the mutation it targets has been shown to fail it. Two
of the six are subtle enough to be worth naming here: a bite that
RAISES is swallowed by the hook's pcall and "passes" for the wrong
reason, and a fixture whose `serve` sleeps cannot reach any ordering
where the primary comes up first.
Closes review round 2 — 1 blocking, 2 high, 1 medium — decides both
remaining open items, and re-integrates canonical `main` @ `ccf29e3`
(#172 + #157; documentation plus one `src/buffer.rs` regression test,
no protocol or Stage 2 source anchor moved).
**R2-1 (blocker) — the seam is three boundaries, not one.** Rev 2 asked
for a single document-bottom accessor. That is wrong: once a panel is
installed the present single value must DIVERGE, because several of its
consumers must not move at all. `text_area_bottom`
(`pmacs-gpu/src/main.rs:8490`) is today `status_band_top`,
`geometry_capacity_bottom`, and `document_text_bottom` at once. Rev 3
defines all three, classifies every one of its ~19 call sites as
status-owned / document-owned / geometry, and records that four sites
rev 2 named (`:3175`, `:3185`, `:6601`, `:6607`) consume a status-band
HEIGHT and no bottom coordinate at all, while the status background
`:5908` and status text `:7134`/`:7922` must stay at the physical
window bottom.
The acceptance is now a contrast assertion: installing a panel moves
every document-owned consumer WHILE the status band stays
pixel-identical. "Everything moved" alone is passed by a blanket
rewrite of the helper, which is exactly the wrong implementation.
**R2-2 (high) — epoch exactness.** `accept_frame_geometry` returns
`Advanced | Duplicate | Rejected` instead of a boolean that cannot
separate reconcile-needed from already-current from stale; if a boolean
is ever kept internally it must be named `advanced`, since `Duplicate`
is also accepted. Rev 2's exhaustion wording permitted retaining stale
geometry, which is not fail-closed — a real resize after exhaustion
would keep painting a panel sized to disowned geometry. The grid path
now clears `frame_geometry` to unknown and reconciles hidden, and the
frontend takes a terminal latch so a retained matching `Present` cannot
resurrect the band; only a fresh session clears it.
**R2-3 (high) — parent acceptance 52 splits.** 2A has no semantic panel
projection, so it can only prove the extracted painter honors an
explicit `None` map plus the `src/window.rs:562` comment fix. The real
contract is production-reachable only in 2B and is reasserted there
beside 42/43/44.
**R2-4 (medium) — touched gates named**: `statusline_segments_acceptance`,
`m11_5_semantic_acceptance`, `gpu_initial_target_acceptance`,
`gpu_font_acceptance`, beside the vterm, folding, and GPU suites.
Open items decided: `BASE_DIVIDER_HEIGHT = 4.0` at scale 1.0, scaled by
`FontMetrics::scale`, whole strip painted `ui.divider` and used as the
exact hover/drag hit rect; `TEXT_TOP` stays `16.0` unscaled, with
Q#BP15a's "all quantities use the frontend's current scale" narrowed to
font-derived metrics and the divider. Wholesale surface-inset/DPI
scaling is recorded as separate work, not smuggled in.
The ledger's bottom-panel lane keeps its census correction and gains
the three-boundary one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Closes review round 1 — 2 blocking, 3 high, 3 revision points — and
rebases the ground truth onto `main` @ `d152120`.
Both blockers were rev 1 asserting something the parent framing
already decided otherwise:
- **R1-1.** Rev 1 said all 23 census reads route through
`primary_document_window`. Q#BP14 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. Rev 1's rule would have broken remote-op validation and
application, `DispatchIdle`, presence, focused search/menu/completion
routing, and terminal bell ownership. §3.2 restores the four classes
as a table and the acceptance asserts each separately — the
focus-class assertions are the load-bearing half, since a test that
only proves "the document is used" passes with them wrongly
rerouted.
- **R1-2.** The three `src/statusline.rs` active reads have two
dispositions, not one. Only `:644` selects the wrong window; `:629`
and `:675` must keep tracking actual focus, because grid contexts
need a truthful `active`, revalidation must notice a focus change,
and parent acceptance 42 requires a document provider to be able to
observe `active = false` while the panel is focused.
The three high findings:
- Q#BP2S1 resolves to frontend-owned epochs (option 1) — a font or
scale transaction can need to invalidate an old `PanelFrame` while
the derived `CellSize` is identical, which daemon value dedup cannot
detect. Rev 2 adds the four-row transition table, splits grid
allocation from semantic acceptance into two APIs rather than one
ambiguous method, moves the grid allocator off `saturating_add` to
checked-with-fail-closed, and defines the initial epoch and both
exhaustion behaviors. Rev 1's "rejects a lower-or-equal epoch
carrying different data" was itself wrong: a lower epoch carrying
identical data is still stale.
- The `panel_capable` flip is narrowed to an authenticated semantic
session negotiated at **v21 or later**. Denying a v20 peer the new
events is insufficient if the daemon still places its window in a
side panel it cannot render — the gate is on placement.
- Parent acceptance criteria 37-55 are declared authoritative and
mapped to slices 2A/2B, with rev 1's eleven drafts demoted to
refinements. The painter-extraction criterion now pins cells, the
returned cursor, the focused window's `view_top` mutation, and
passive-window state.
All four scout obligations are closed (§5), and the pixel formula is
treated as contract work, not implementation detail:
- The shared/terminal-only validator boundary is named exactly.
- Four new outbox tail-coalescing tags beside the existing four.
- **`State::mono_advance` is unsafe to adopt**: absent a `FontFacts`
probe it samples the document's first shaped glyph, which would make
panel columns document-dependent. The declaration uses the existing
stable normal-face `probe_mono_advance` instead, and declares zero
usable geometry when it returns `None`.
- `BASE_DIVIDER_HEIGHT` does not exist. Rev 2 decides its scaling and
requires **one** document-bottom accessor routing every consumer
(caret, hits, minimap, terminal geometry, clipping, edge scrolling)
— a second unrouted seam is precisely the Stage 1 `Layout::compute`
two-caller defect. The concrete base value is left open for round 2.
Also: the coherence statement now names journey steps 7-10 instead of
claiming none, and drops rev 1's overclaim that this advances
background-work visibility — a panel gives output a placement but adds
no join key to COHERENCE §9's four disjoint activity planes.
The ledger's bottom-panel lane is updated from "no branch and no
framing yet" to the framing's real state, and carries an explicit
correction: that entry was itself the source of rev 1's census
mis-statement.
Factual corrections: `InitialTargetResult` is at `message.rs:1145`;
`primary_document_window` has four references and two production paths
(`daemon.rs:1639`, and `daemon.rs:2998` via `primary_document_buffer`,
which is census #22); fifteen PRs merged since the parent's last
re-scout, not eleven.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The re-framing `docs/bottom-panel-framing.md` rev 4 §2 requires before
Stage 2 (the GPU panel band) is implemented. Re-scouted against
canonical `main` @ `5aa9044`, protocol v20.
It does not restate the parent's decisions; it records what the
re-scout found. Every source anchor Stage 2 inherits had moved, but
none of the parent's mechanical model was falsified. Two facts held
and are load-bearing: protocol is still v20, so Q#BP9 resolves to
**v21** with no reservation needed, and both byte pins
(`InstanceMessage::InitialTargetResult`,
`FrontendEvent::TerminalPointer`) are still their enums' final
variants.
Four findings:
- **Q#BP2S1, new and open.** Stage 1 landed a daemon-side geometry
epoch allocator (`declare_frame_geometry`), but Q#BP15a specifies a
frontend-owned epoch echoed by every `Present`. The landed allocator
also dedups on value and uses `saturating_add`, which is neither
wrapping nor the fail-closed the framing asks for. Three resolutions
are stated with a recommendation.
- **The §1.3 census is essentially unrouted.** Stage 1 built the
`primary_document_window` seam but it has one production caller;
~80 direct `.active` reads remain. This is Stage 2's bulk, not its
tidy-up, and the stage plan sequences it first.
- **The statusline active read is three sites, not one.**
- **Four scout obligations are still open** and are named rather than
papered over, including the GPU-side pixel formula inputs.
Also carries the staged plan, draft acceptance criteria, the coherence
impact per `COHERENCE.md` §20 (§14 is the section it serves), and four
questions for the user.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Stage 1 of the terminal config/copy-mode arc is in review; Stage 2 is
not started. Records the four decisions forced by scouted ground truth,
the four bites against four different wrong implementations, the two
reusable test instruments, and the gate results.
Stage 1 of the terminal config/copy-mode arc. The terminal had no
configuration surface at all: the command hardcoded $SHELL, scrollback
was a per-open argument only, and the escape chord was a literal in
Rust. No protocol change.
Profiles are a raw Lua table, pmacs.terminal.profiles, not a registry
setting: ConfigValue is four scalars with no table kind, so profiles
join pmacs.lsp.config and pmacs.pair.sets until table-valued settings
exist. The registry gains three scalars whose defaults reproduce the
previous behavior exactly.
Field resolution is explicit open argument, then profile field, then
scalar setting, then $SHELL. env MERGES, with explicit entries
overriding the profile's, because first-wins there would silently drop
half a user's environment. An explicitly named profile that does not
exist is an error even when terminal.default-profile is valid, so a typo
cannot silently fall back.
The two open-time settings resolve through the GLOBAL chain, because
they are read before the identity buffer exists and no caller could have
pinned a local override on a buffer that does not yet exist. Only
terminal.escape-key resolves per buffer, which makes a per-terminal
escape a supported feature.
The escape key is parsed at most once per (terminal, config epoch), and
the cache lives on TerminalSession so its lifetime is the terminal's,
with no purge hook to forget. The epoch alone is not a sufficient key:
it does not advance when focus moves between two terminals with
different buffer-local values, so an epoch-only cache serves one
terminal's chord to the other. An unparseable value falls back to C-c
and reports once per terminal per effective invalid value through the
status line, because a terminal with no escape chord cannot be escaped
to fix the setting that broke it.
Repeating the escape now sends THAT chord to the child through the
ordinary key encoder, rather than a hardcoded ETX. With an escape of
C-x, the previous code sent Ctrl-C and made literal Ctrl-X unreachable.
C-c t opens a terminal. COHERENCE Priority 1 names a terminal
keybinding, and section 2 step 8 grades the terminal works-but-
undiscoverable; C-c is already a live global prefix, so this is a new
leaf rather than a shadow. It is unreachable from inside a terminal,
where C-c is the escape.
Acceptance is tests/terminal_config_acceptance.rs, deliberately NOT
crdt-gated so CI actually runs it. Four bites, each against a different
plausible wrong implementation: a hardcoded ETX fails acc6/9; an
epoch-only cache key fails acc7; a single last-entry cache fails acc8's
parse count; removing the invalid-value fallback fails acc10.
Two test-instrument notes worth keeping. cat -v is the echo probe
because the screen rejects C0 controls before they reach cells, so a raw
echoed Ctrl-X would be invisible. And the probe counts occurrences
rather than testing presence, because a single-character probe collides
with the child's own banner text.
Three P1 lifecycle defects and two P2s. The focused suite was 20/20 with
every one of them live, which is the part worth keeping.
**1. The crashed primary respawned forever underneath the fallback.**
Round 2 skipped the retire call for terminal servers to avoid corrupting
them — but the crash had already armed `next_restart_at`, and
`maybe_restart` fires on every elapsed backoff with no attempt ceiling.
The broken command kept respawning under the live fallback.
The right call depends on the state, and each is wrong for the other:
`forget` REQUIRES a terminal state and removes the client outright,
which also drops the restart timer; `stop` is for a live one and
corrupts a terminal one (its not-initialized branch parks it in
`ShuttingDown` forever). `retire_server` now dispatches on state.
**2. Re-attachment targeted whatever buffer was active when the
asynchronous verdict landed.** `_attach_buffer` is an active-buffer-only
seam, and "some attachment now names a different server" is satisfied by
an unrelated Rust buffer — clearing the retry and leaving the Lean buffer
stale forever. The initiating buffer is now captured and the retry waits
for it.
**3. A failing fallback retried every tick forever, silently**,
contradicting acceptance 27's promise that a second failure surfaces.
"Waiting for the old server to go" and "attempting the replacement" are
now separate: once the old one is terminal or gone, the replacement is
attempted EXACTLY once, and a spawn failure is reported.
**4. The Lake version parser was being applied to arbitrary wrappers.**
`version_below_3_1` encodes lake's output contract; a working
`my-lean-wrapper` reporting "wrapper 1.0" would have been replaced
despite its server initializing fine. The version probe is now gated on
the command's basename being `lake`. The FAILURE latch stays
command-agnostic — that one keys on the server actually not starting,
which is true of any command.
**5. An unconfigured Lean server was reported as a failure** and latched,
poisoning the session so a later configuration could never take effect.
Absent config or command now means disabled; only a configured command
that produced no attachment is a failure.
**6. The ledger recorded pre-fix counts** after the fixes were pushed.
Now 25/25 and 3,214. That is the #161 fmt-blocker error in a slower
form: verification must describe the pushed tree.
Sign-offs requested in review: `M.fallback` is now `M._fallback`, an
underscored test seam, and its idempotence check compares args as well as
command — the same command with different arguments is not "already
applied". Dropping the `command ~= "lake"` guard stands for the failure
latch only.
Five regression tests added, and **three of them were too weak on first
write; only bite-testing found it**:
* asserting "no live non-fallback server" misses a respawn loop,
because a respawning server sits in `crashed` most of the time —
`attempt` is the observable that counts respawns;
* returning to a buffer with `find_or_open` re-fires
`buffer.after-load`, which repairs the attachment regardless of the
code under test — `switch_buffer` is the honest return;
* a MISSING command fails synchronously inside `after-load` where the
rebuild happens inline, so the async race cannot occur — only the
probe path exercises it.
Each of the five now fails against the exact round-2 mutation it targets.
Brings the three required docs current after #158 merged, and discharges
the follow-up that framing named for itself.
COHERENCE.md section 16 audits the claim that the GPU frontend exceeds
the TUI "under real divergence pressure" without a privileged frontend
emerging. Inline math is the sharpest instance of that so far -- the GPU
typesets $...$ while the TUI shows LaTeX source, and the TUI fallback is
a named deferral. Section 25 makes that update ride the PR, so the
enumerated list gains the case along with what keeps it inside the rule:
the slice reserves no protocol version and adds no wire surface, so the
divergence is presentational and both frontends read the same model.
docs/inline-math-framing.md carried a licence error the slice framing
flagged in its own section 9 and deliberately did not fix in-branch,
since the parent is a merged document. Latin Modern Math is under the
GUST Font License, not the OFL; the row now says so and records the
~717 KiB bundled size.
docs/agent-handoff.md records the landing and re-anchors section 1 to
d152120. The bullet leads with the facts a fresh agent would otherwise
have to rediscover: the whole slice lives in pmacs-gpu because pmacs-gpu
depends only on pmacs-protocol and never on pmacs; the v0 subset is 34
Greek symbols, sub/superscript and \frac; an unsupported command fails
the WHOLE span back to source, so most inline spans in a real paper
still show LaTeX by design; and math is suppressed while the caret is
inside its span.
docs/active-work.md removes the merged lane per its own update protocol
and adds a Closed entry. Four things there are reusable beyond this arc:
a stale frontend binary is invisible from the source tree, so diagnose
with strings on the binary rather than by re-reading a checkout that is
already current; the dangerous integration was the one that did NOT
conflict, so decide from the shared-file set rather than from whether
git complained; integration is proved by predicting the other side's
test-count delta and checking it; and m4_5_basedpyright has no timeout,
hangs forever, and is intermittent, so an earlier clean sweep proves
nothing. It also corrects a claim I recorded on main: the branch's
missing CI was not an unidentified cause -- a conflicting PR builds no
merge ref, so no pull_request run is created.
Docs only; no code changes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HZjWMjwPXhPbt9upku9mCk
Round 1 review, four P1s. All real; the first two mean the fallback did
not work at all.
**1. The latch swapped the config but never spawned or re-attached.**
Nothing re-fires an attach on a config change and `attach_buffer`
early-returns for a live attachment, so the buffer stayed bound to the
server that had just been stopped. The user got a config edit and no
language server. `fire_latch` now rebuilds through a new
`pmacs.lsp._attach_buffer` export.
Two mechanics had to be right for that rebuild to happen at all:
* It is **retried on the tick**, because `pmacs.lsp.stop` leaves the
state `shutting-down`, which `server_is_live` counts as LIVE — an
inline re-attach early-returns the stale record and the swap is a
silent no-op.
* The latch **does not stop an already-terminal server**, and this is
a substrate bug worked around rather than a style choice.
`LspManager::stop` on a `Crashed` client takes its not-initialized
branch, terminates the dead process, and sets `ShuttingDown { ..
None }` on the premise that "the next exit observation cleans up" —
but the exit already happened, which is what made it `Crashed`. No
further event arrives, so the client is stuck in `ShuttingDown`
forever: `server_is_live` reads it as live so `attach_buffer` never
rebuilds, and `forget` refuses it for not being terminal. Stopping a
dead server is what makes it un-replaceable. Named in framing §6; the
fix belongs in `stop` and changes behavior for every language.
**2. A missing `lake` bypassed probe and latch entirely** — the single
most likely real failure. `ensure_server` swallows a synchronous ENOENT
and returns nil, so there was no attachment, and the hook keyed on
`active_attachment()` returned before arming anything. The hook now keys
on the buffer's LANGUAGE and treats a Lean buffer with no attachment as
the failure itself.
**3. `waitForDiagnostics` omitted `version`.** Lean's
`WaitForDiagnosticsParams` is `{ uri, version }` (v4.9.0,
`src/Lean/Data/Lsp/Extra.lean`); the request is how a client says which
revision it wants. It looked correct only because the fake server echoes
any payload — so the fake server now validates and returns InvalidParams
without it.
**4. The ledger stated the dangerous stacking order** in one sentence
and the correct rule in the next. Fixed to say BEFORE. A safety rule
written twice with opposite senses is worse than not written.
Also (P2): the probe/latch suite now drives the production path —
`buffer.after-load` -> ticks -> probe drain -> latch -> re-attach — with
real executable stubs, and asserts the originally opened buffer ends up
on a LIVE server. Round 1's acceptance 36 asserted every server was
terminal, i.e. pinned the ABSENCE of the fallback it claimed to test.
`M.fallback` is a table so the suite can point it at a working stand-in;
the probe now spawns `cfg.command --version` rather than a hardcoded
`lake`, which is also more correct for a user who configured a wrapper.
`swap_to_fallback`'s `command ~= "lake"` guard is gone: the latch fires
only when the configured server actually failed, one visible fallback
beats no server, and `probe.latched` is what keeps it to exactly one.
Three new bites, all against the committed tree: no re-attach after the
swap -> three latch tests fail; hook keyed on the attachment -> the
missing-`lake` case fails; `waitForDiagnostics` without `version` ->
acc37 fails with the server's InvalidParams.
Two stages, one arc, no protocol change. Stage 1 makes the terminal
configurable (profiles, scrollback, escape key) and binds the opening
command; Stage 2 adds copy mode and search over scrollback. They are
independently releasable and get separate branches and PRs.
Three scouted facts shaped the design, two of them ruling out the
obvious plan.
Profiles cannot be a config-registry setting: ConfigValue is four
scalars and there is no table kind, so profiles join pmacs.lsp.config
and pmacs.pair.sets as a raw Lua table while the registry holds only
scalars.
Search cannot reuse isearch in place: SearchStore addresses matches as
byte ranges into a buffer's rope, and a terminal identity buffer is
empty by construction.
An in-place copy mode would be the seventh dispatch shadow, which
COHERENCE section 6 grades weak and growing by one island per modal
feature, with no transient-keymap mechanism to migrate to.
Copy mode therefore materializes the retained rows into an ordinary
read-only buffer. isearch, motion, selection and the kill ring work
with no new substrate; the "keys must not reach the child" problem
dissolves because the snapshot is not a terminal; and describe-key
stays truthful because the bindings are buffer-local. The cost, stated
in the doc, is that the snapshot is point-in-time rather than a live
freeze.
Four review rounds produced the load-bearing parts: the escape-key
cache is owned by TerminalSession so its lifecycle is the terminal's,
with three acceptance pins that each fail a different wrong cache; the
snapshot needs set_round_trip_input because a Lua intercept does not
set Buffer::read_only and an optimistic CrdtOp would mutate both the
daemon buffer and the mirror; the double-escape must encode the
configured chord rather than a hardcoded ETX; and the two open-time
settings resolve through the global chain because they are read before
the terminal buffer exists.
No code changes in this commit.
`main` moved through #158-#166 (Lean 4 Stage 2, COHERENCE.md, find-file,
the dired framing and Stage 1, the GPU terminal input fix) while this
documentation branch waited. Both required docs conflicted; neither
conflict was a code signal.
Resolution:
- `docs/active-work.md`: main's ledger is the base — every lane it has
gained since this branch was cut is kept verbatim. Only the
bottom-panel lane is replaced with this branch's "Stage 1 MERGED;
Stage 2 (GPU band) is next" section, and only the bottom-panel entry
is added to "Closed since the last snapshot".
- `docs/agent-handoff.md`: main's version is the base. This branch's §1
bottom-panel bullet, its §1 roadmap Arc 7 entry (which also records
that DAP is now unblocked), and its four §5 ops lessons are inserted
at their anchors.
One repair rides along. Main's `docs/agent-handoff.md` carried a
garbled fragment at §1: a duplicated, truncated "GPU initial target
LANDED — #148" bullet whose body was the tail of the old head-of-`main`
anchor bullet, leaving the `SUPPORTED=[6..=20]` protocol enumeration
orphaned mid-sentence. The fragment is removed and the enumeration is
restored as its own bullet.
No code changes; the merged tree's non-doc content is main's.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Framing Q#LN7, Q#LN8, Q#LN16; acceptance 22–28, 24a/24b, 35, 36, 36a, 37.
Stacked on Stage 3a (#167), whose notification/response seams and
`pmacs.fs.canonicalize` this consumes. No protocol change; the only Rust
outside the test helper is one `include_str!` line.
**The Lake-aware root (Q#LN8).** `pmacs.project.detect` cannot express
this rule — it is innermost-wins by construction, and a Lake package's
`lean-toolchain` sits at the outermost level, so a file under
`<pkg>/.lake/packages/dep/` belongs to `<pkg>`'s server rather than
`dep`'s. The resolver walks up collecting markers and returns the
outermost, stopping at `pmacs.project.search_boundary()` so a stray
marker above a fixture cannot leak in.
Two things about the marker test are easy to get wrong in opposite
directions, and both are pinned. `io.open` **succeeds on a directory**,
so a truthiness check accepts a `lean-toolchain` directory; but
requiring a non-nil read rejects an **empty** `lean-toolchain`, which is
a legitimate marker — `locate-dominating-file` semantics are existence,
not content. The discriminator is `read`'s second return: decline only
on a non-nil error. Acceptance 24a and 24b each fail against the
implementation that satisfies only the other; both bites are recorded.
The root is canonicalized once up front, because a configured root
reaches `file_uri_for` verbatim and that URI is the affinity key (#161).
Canonicalizing the starting directory suffices — every ancestor of a
canonical path is canonical, since the walk only strips components.
**`lake serve` with a lazy probe and a one-shot latch (Q#LN7).** Nothing
runs at init: `pmacs.lsp.config` is declarative, and spawning a process
at startup for every user, Lean-using or not, is the cost rev 1 refused.
Both the probe and the server spawn are gated on a real Lean attachment.
The probe cannot gate the first attach — there is no blocking process
run, so its verdict arrives after `ensure_server` has already decided.
Hence the optimistic spawn, with the probe and latch correcting it. A
non-zero probe exit is deliberately NOT a trigger: §2.9's elan-shim case
makes `lake --version` fail on machines where `lake serve` still works,
and the server-failure latch covers that better. The probe answers only
the question failure detection would answer slowly — an old-but-working
lake that starts a useless server.
The latch stops the failing server **before** spawning the fallback, and
that ordering is load-bearing rather than defensive: the spec default is
`OnCrash`, the termination handler never consults the exit code, and
`maybe_restart` has no attempt ceiling, so a broken `lake` respawns
forever underneath the latch. `pmacs.lsp.stop` sets `restart = Never`,
which is what disarms it. Bitten: removing the stop fails acceptance 36.
The swap rewrites `command` and `args` only, so a user's `env`,
`settings`, `init_options` and `root` survive — a wholesale table
replacement would discard their `init.lua` at the moment they are least
likely to notice.
**`waitForDiagnostics` (Q#LN16)** resolves through Stage 3a's response
seam, with `M-x lean.wait-for-diagnostics` on top. `$/lean/fileProgress`
subscribes on the notification seam and is pinned end-to-end through a
new `leanprogress` mode on the fake server rather than by calling the
handler directly — the wiring is the only part that can break.
**Attribution (COHERENCE §9/§1.2).** The probe spawns as
`lean:lake-version-probe`, so a user wondering why their editor touched
`lake` finds an owner in `pmacs.process.list`. The latch reports through
`pmacs.editor.set_status` — the channel that exists — and acceptance 36a
observes that channel, so a report made only through the undefined
`pmacs.error` would fail it.
**Stage 1's acceptance 12 is updated, half superseded.** It asserted
`pmacs.lsp.config.lean4 == nil` to guard against a Stage-3 front-run;
Stage 3b is that stage, so keeping it would pin the opposite of the
intended behavior. The half that survives is the one about restraint,
and it matters more now: constructing an editor spawns nothing even
though the config exists and names `lake`, and opening a Lean buffer
with no server configured spawns nothing either. That is what holds
Q#LN7's "not at init" promise.
Bites recorded, all against the committed tree: bare `io.open` -> 24a
fails, 24b passes; require-non-nil-read -> 24b fails, 24a passes; no
canonicalization -> the symlink case spawns two servers; no stop before
fallback -> acceptance 36 fails.
Third merge of main into the lane, at b889873 (GPU terminal input #166).
Unlike the first two this one produced NO conflict -- and it is the case
that shows why a clean git merge-tree is not a reason to skip
integrating. #166 lands 41 lines in pmacs-gpu/src/main.rs, the same
heavily-rewritten file as the first integration; the two edits merged
silently only because they sit in different regions of it (#166 is
entirely in the headless probe, this lane rewrites the render path).
Merging the PR on that clean auto-merge would have shipped a combination
no gate had run.
Reconciliation, run against what #166 actually added rather than against
a pass/fail: it adds 3 library tests, 2 to vterm_stage3_acceptance, and
0 to pmacs-gpu. Predicted lib 1,826 -> 1,829, CRDT 2,003 -> 2,006, GPU
unchanged at 202; that is exactly what ran. Suite count 91 -> 92 is
#161's new lsp_multi_root_acceptance binary. All three sides' markers
verified live in the shared file.
Also records an ops trap that cost hours this session:
m4_5_basedpyright_initializes_and_negotiates_encoding does not time out,
it hangs forever, parking a --workspace sweep at 38 of 92 suites with a
live basedpyright langserver child. The per-suite M4 gate already skips
it; the workspace sweep needs the same flag.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HZjWMjwPXhPbt9upku9mCk
CI round 1: both macOS jobs failed on the acceptance case added last
commit. APFS enforces valid UTF-8 in filenames, so `std::fs::write` with
a 0xFF byte in the name fails with EILSEQ ("Illegal byte sequence")
before `pmacs.fs.canonicalize` is ever called. The fixture cannot be
built there.
That is a filesystem refusing to represent the case, not a behavioral
difference: the subject — `to_str()` returning None for a non-UTF-8
resolution — is platform-independent Rust, and the Linux run pins it.
`#[cfg(unix)]` was the wrong granularity; review had asked for unix
gating on the symlink tests and I applied the same gate here without
checking whether the filesystem, rather than the API, was the
constraint.
Gated `#[cfg(target_os = "linux")]` with the reason in place, rather
than skipped at runtime, so a future failure here is a real failure and
not a silent no-op.
Ledger records both CI-round facts: this one, and that
`composition_overhead_under_ten_percent` is load-sensitive under a
parallel workspace sweep (it reported -4.6% realistic overhead in the
same run that tripped its 10% budget at 18.8%, which is noise, not work).
Six findings from review, one of them a real defect.
**`canonicalize` could emit a path that exists nowhere.**
`p.display().to_string()` substitutes U+FFFD for non-UTF-8 bytes, so a
resolution landing on such a path returned a plausible-looking string
that does not exist on disk — worse than nil, because this value becomes
a server-affinity key through `file_uri_for` and would silently fail to
round-trip, while the doc promised nil for anything unresolvable. Now
`.and_then(|p| p.to_str().map(str::to_owned))`: unrepresentable is a
decline, matching how the fs layer already treats non-UTF-8 symlink
targets.
Pinned by a new acceptance case that reaches a non-UTF-8 target through
an **ASCII** symlink, so the input is representable and only the
resolved output is not — the case a UTF-8 check on the argument would
miss. Bitten: restoring `display()` fails it.
The other five:
- `on_response`'s doc comment now warns that registering against a
server with no attached buffer is fire-on-death, not fire-on-reply,
because the drain visits only attached sids. It looks exactly like a
hung request while debugging, and 3b is the first caller likely to
hit it.
- `deliver_response`'s comment still carried the pre-correction
rationale ("removed BEFORE invocation ... must not be re-entered") —
the claim the bite disproved. It now says what is true: removal is
unconditional, before-vs-after is unobservable without a re-entrant
drain, and the reachable bug is gating removal on a clean return.
- Dropped `server_attempt`'s unused second return.
- Deleted a vacuous assertion in the no-attachment test (counting `_G`
entries to assert "lua globals are readable") — scaffolding that
pinned nothing, the exact shape the project's own lesson flags.
- `#[cfg(unix)]` on the three symlink-dependent tests.
Gates re-run in full. The sweep's first pass tripped
`composition_overhead_under_ten_percent` at 18.8% against a 10% budget;
it passes 3/3 in isolation here, passes in isolation on main, and the
same run reported the realistic-frame overhead as **-4.6%** — a negative
figure is measurement noise, not added work. Nothing in this diff is on
the render path. Rerun of the full sweep: 3,189 across 93 suites, zero
failures.
Main advanced twice inside one review round (#161, then #166), the
second landing while the first integration's sweep was still running.
The ledger now names both integrations, how each doc conflict was
resolved, and the verification numbers for the twice-merged tree -- plus
the lesson that a lane in review against a fast-moving main reruns its
gates per integration, not per push.
Main moved again while this lane was gating: the GPU terminal-input fix
merged as #166. One conflict, in COHERENCE.md's journey table, resolved
as the union -- this lane owns step 7's file half, #166 owns step 8's
GPU-terminal addendum.
The module doc said an uncaught raise inside a `pmacs.async` coroutine
"goes to *errors*, not the status line". #161's COHERENCE finding shows
that is wrong, and in the worse direction: `pmacs.error` is never
defined in production, so `step()`'s guarded report is dead and the raise
falls through to a bare `error()` inside `pmacs._async.tick()` -- whose
result `EditorState::tick_async` discards with `let _ =`. The failure
reaches nowhere at all, and dired would look like it silently did
nothing.
So the per-coroutine `pcall` plus `pmacs.editor.set_status` is
load-bearing, not tidy, and the doc now says which channel is dead, which
is live, and that the acceptance suite observes the live one -- the
corollary COHERENCE draws from that finding.
The ledger records the integration, the reruns on the merged tree, and
the ops lesson that cost three CI runs: a conflicting PR has no merge
ref, so GitHub creates no `pull_request` run and nothing reports the
absence.
Multi-root LSP affinity merged as #161 (`main` @ `46a1b8f`) while this
lane was in review, which made the PR conflict -- and a conflicting PR
has no merge ref, so GitHub silently stopped running CI on it after the
first push. Integrating rather than rebasing, per the #135/#137
precedent: the review anchors stay addressable and every gate is rerun
against the merged tree.
One conflict, in COHERENCE.md's in-flight list, resolved as the union of
both truths -- and #161 is now merged, which its own text still called a
PR.
The overlap to watch is `src/lua_bindings/mod.rs`: #161 widened the
`lsp.list()` row builder while this lane added `pmacs.path` and the
read_dir listing conversion. The merge was textually clean, which the
folding arc's lesson says is not the same as compiling, so the full gate
suite reruns from here.
Rev 5 said acceptance 34's second edge was a killed buffer. Implementing
it showed that is false: the Rust core fires exactly five hooks —
buffer.after-edit, buffer.after-load, buffer.after-switch,
frontend.detached, process.after-tick — and there is **no buffer-kill
hook**, so lsp.lua never tears an attachment down and the drain keeps
reaching that server. The premise (the drain builds its sid list from
`attachments`) was right; the inference needed attachments to be removed
on kill, and nothing removes them.
The reachable leak has the same root cause by a different path.
`attach_buffer` drops a sid from `attachments` the moment
`server_is_live` reports false and rebuilds against a fresh server — so
`crashed` / `stopped` is the event *least* likely to be drained, and an
event-driven purge leaks in exactly the case it exists for. The purge
therefore polls `pmacs.lsp.list()`, which enumerates the manager
directly. Acceptance 34's second half now exercises a server in **no**
attachment, which is the shape that discriminates: bitten, an
event-driven purge fails it while the attached case still passes.
§0.1 finding 6, Q#LN9, and acceptance 34 all updated; the wrong wording
is left visible with its correction rather than quietly replaced, since
the mistake is the useful part.
Ledger gains the Stage 3a lane: branch, worktree, what ships, both
corrected claims, the `install_async` load-order trap, the recorded
bites, the one knowingly unpinned guard, and gate results.
Two corrections found by bite-testing the suite rather than by reading
it.
**Acceptance 32 was mis-named.** It claimed to pin "the one-shot is
removed BEFORE invocation". Biting that — moving the removal after the
`pcall` — still passes, because `pcall` catches the raise either way and
the removal runs regardless. The before/after ordering is unobservable
unless a handler re-enters the drain, and nothing does. What the test
actually pins is that removal is **unconditional**: the bite that gates
it on `if ok then` fails 2 != 1, because the surviving registration is
invoked a second time by the purge. Renamed and re-commented to say so.
The implementation still removes before invoking, which is the right
defensive order; it is simply not what the assertion proves.
**The purge's generation check is defensive and untested**, now labelled
in place instead of reading as covered. Reaching it needs a crash and
its restart to both land in a gap with no `_async.tick`; the backoff is
500ms, so any tick in that window sees `crashed` and the
absent-or-terminal test fires first. Every attempt to stage it
deterministically exercised the `crashed` path instead. It stays as
insurance for a stalled editor, and says that about itself.
Bites recorded, all against the committed tree:
- removal gated on a clean return -> acc32 fails (2 != 1).
- purge driven by a `crashed`/`stopped` event seen in the drain, the
design the framing originally implied -> the no-attachment case fails
("never called"), while the attached case still passes. That is the
discrimination acceptance 34's second half exists for.
- a resolver without `pmacs.fs.canonicalize` -> two servers, pinned as
34b's own falsification.
Arc 8 Stage 3a (framing Q#LN9, Q#LN20). No Lean content: this changes
the event drain every LSP language runs through, and is split from the
Lean server work for the reason Stage 2 was.
**The seams.** `handle_server_requests` handled five `request` methods
and `initialized`, dropping every `notification` and `response` on the
floor. Dropping responses made `pmacs.lsp.send_request` a write-only API
from Lua — the reply was drained and discarded, so nothing outside
Rust's typed stores could consume one. Two new arms route to
`pmacs.lsp.on_notification(method, fn)` (persistent, method-keyed) and
`pmacs.lsp.on_response(sid, request_id, fn)` (one-shot). Both extend the
existing loop rather than opening a second `events_take` caller, which
would steal events from it.
A one-shot is removed **before** invocation, so a raising handler cannot
be re-entered. Every subscriber is `pcall`ed and a raise reports through
`pmacs.editor.set_status` per COHERENCE §1.2 — not `pmacs.error`, which
is defined nowhere in production. The notification list's length is
captured before the walk so a subscriber registering another cannot
extend the list being iterated.
**The purge is driven off `pmacs.lsp.list()`, not off a death event.**
The framing said acceptance 34's second edge was a killed buffer. That
was wrong, and scouting the implementation is what caught it: pmacs
fires exactly five hooks (`buffer.after-edit`, `buffer.after-load`,
`buffer.after-switch`, `frontend.detached`, `process.after-tick`) and
there is no buffer-kill hook at all, so `lsp.lua` never tears an
attachment down and the drain keeps reaching that server. No leak there.
The real leak is a different path with the same root cause. The drain
builds its sid list from `attachments`, and `attach_buffer` drops a sid
from that table the moment `server_is_live` reports false — rebuilding
against a fresh server. So the `crashed` / `stopped` event that should
trigger the purge is precisely the one most likely to go undrained. A
purge wired to that event leaks exactly when it matters.
`pmacs.lsp.list()` enumerates the manager directly and is unaffected by
attachment bookkeeping, so the purge polls it after each drain: a sid
that is absent, terminal, or running a **new generation** settles its
pending one-shots with an error. The generation check uses the `attempt`
field, because a crash-then-restart reuses the sid — without it a
one-shot would sit waiting on a reply the dead generation owed.
**`pmacs.fs.canonicalize`** (Q#LN20) is the one synchronous function on
`pmacs.fs`, and synchronous is the point: its consumer is a
function-valued `config.root` called from `ensure_server` <-
`attach_buffer` <- `buffer.after-load`, where there is no coroutine and
`pmacs.fs.stat`'s awaitable handle is unusable. It is installed from
`install_async` rather than `install_project` purely for load order —
`make_workspace` runs after `fs.lua` is evaluated, so a canonicalizer
placed there reads nil.
Acceptance: `tests/lsp_dispatch_seams_acceptance.rs`, 14 tests, driven
against `pmacs_fake_lsp` through rust so nothing needs a toolchain.
Dispatch integrity is exercised at real co-occurrence — the fake server
writes `workspace/applyEdit` and the `executeCommand` reply back to
back, so both land in one `events_take` batch. 34b asserts affinity
survives a symlinked open and is paired with its own falsification: the
same resolver minus the canonicalize call spawns two servers, so the
positive case cannot be vacuous.
COHERENCE.md section 25 and the handoff/ledger update protocols make
these ride the PR.
COHERENCE.md:
- Section 6 named one optimistic key classifier and attributed it to the
GPU. There are two, one per replica frontend:
crate::optimistic::classify_key belongs to the pmacs --attach TUI
replica, and pmacs-gpu has its own unrelated optimistic_insert_text /
optimistic_crdt_insert. The section's "kept honest by
dispatch_idle_for" claim is confirmed for both, which this
investigation verified rather than assumed.
- Section 16 graded per-frontend degradation strong on the evidence of
per-frontend fold projection. That grade stands, but the practice is
enforced by convention rather than structure, and this defect is the
counter-example; the note says so and points at what is now structural.
- Section 2 step 8 records that the terminal was broken outright on the
GPU frontend, not merely undiscoverable.
docs/agent-handoff.md section 5 gains four lessons: adjacency does not
make two operations alternatives (and two individually sound idempotence
guards can be jointly useless); bite against every pre-image the fix
could have taken, since the obvious guard here fixes the storm and
introduces a controller leak; a quiet child is an instrument, because a
frame storm hides inside a chatty fixture and a geometric readout is
satisfied by an oscillating geometry; and TerminalMode::Raw makes
sh-based input fixtures useless because there is no ICRNL.
docs/active-work.md gains the lane entry with the branch, the bite
matrix, the named out-of-scope items, and the gate results.
Framing rev 7 adds S1-10..S1-12 -- the three findings that changed
behavior, each stated as the durable lesson rather than as a diff:
painting takes a buffer and seating takes the world, so any post-await
cursor operation needs an active-buffer guard; the rendered columns are
a contract Stage 3 is planned against, so precision yields to width; and
`open_directory`'s changed-nothing-on-failure invariant is itself a
probe, which is why the symlink descent no longer lists the target
twice. Plus the tolerant-channel note: cancellation was never a backstop
for a dired listing, because nothing cancels one.
The ledger records the round, the updated counts (dired 25 + 25 CRDT,
sweep 3,189 across 92), and the process lesson that cost me the fixes
once: a mutation-bite helper restores with `git checkout --`, so a fix
must be committed before it is bitten.
The dispatcher loop applied BOTH terminal-layout syncs to EVERY attached
frontend. A semantic session satisfies both conditions, because it has a
term_sizes entry from AttachRequest and a semantic terminal declaration,
so its PTY was resized twice on every tick forever: the grid arm
installed the TUI placement size, the semantic arm installed the declared
content rectangle, and each arm's own idempotence guard only ever saw the
size the other had just written. The child took a SIGWINCH storm at tick
cadence and the screen reflowed continuously, which is what made typing
into a GPU terminal impossible while output kept flowing.
The grid arm is also the only per-tick controller-liveness release a
semantic frontend gets, so simply skipping it for those frontends trades
one defect for another: a GPU window that switches away from its terminal
would hold the controller forever, and no peer could resize that PTY
again. The semantic arm cannot take over that job, because the
buffer-follow snapshot clears the viewport declaration that would drive
it.
sync_terminal_layout is therefore split into a frontend-kind-neutral half
(panel reconciliation plus controller liveness, which read only views,
windows and the controller) and a grid-only geometry half (TUI placement
plus the resize). The dispatcher runs the neutral half for every attached
frontend once per tick, then exactly one geometry arm per frontend kind.
sync_terminal_layout survives as the composition of both halves, so the
in-process editor loop and LOCAL are unchanged.
The loop body is extracted into sync_terminal_layouts_for_tick, which
makes the grid/semantic exclusivity structural rather than two adjacent
ifs, and lets the tests drive the real loop body instead of a
re-implementation.
The release that fires when a window has no placement stays in the grid
half deliberately: a semantic frontend has no window_placements entry at
all, so moving it into the neutral half would release a GPU session's
controller on every tick.
Bite-verified against two pre-images, because one is not enough here --
the naive guard fixes the storm and introduces the controller leak, so a
single revert would score the fix complete when it is not:
pin main naive guard split
settle (acc 2+3) FAIL pass pass
controller release (acc 6) pass FAIL pass
grid still resizes (acc 5) pass pass pass
Real-path acceptance: a quiet child that counts SIGWINCH reports 144
frames in 4 s and WINCH 1..12 on screen against the pre-fix tree, versus
a settled screen with the fix. Acceptance 4 (input reaches the child and
returns) is a keep-working pin and passes on both sides -- key transport
was never the defect.
No protocol change; stays v20.
Levi Neuwirth