Implements docs/lsp-file-watch-d3-framing.md revision 4, approved
2026-08-11 with the four rulings adopted as proposed: the honest bar
(absent at idle, one attributable job per concurrently due group), no
exclusions by default, server root_uri -> cwd -> attachment fallback,
and constants rather than config keys.
pmacs.fs.walk_tree: the whole recursive tree as ONE cancellable job
(JobKind::FsWalkTree, reply reuses ReplyKind::ReadDir --- identical
payload shape, so the Lua boundary needs no second conversion). Names
are base-relative; symlinks recorded, never traversed; an unreadable
subdirectory skips its subtree (scan_tree's pcall behaviour); only the
root failing to open fails the walk; the cancel token is polled once
per directory. Eight Rust unit tests, including flat-directory entry
parity with read_dir_blocking and the two review-round cancellation
cases (empty-tree pre-cancel; mid-walk via the cfg(test) entry hook).
The watcher itself is rewritten as the framed group scheduler. No
sleeps anywhere: one process.after-tick subscription (installed once
and guarded --- pmacs.hook.remove does not exist) drives every
(server, base) group's deadline off monotonic_ms, autosave's Q#AS2
idiom. The old design held one pool thread per sleeping watcher and
allocated 1 sleep + D read_dir jobs per watcher per tick --- 1,326
per tick for rust-analyzer's six watchers on this 220-directory
checkout. At idle there is now NO running job, which is also the
strongest witness in the suite: activity_summary settles to None, and
that assertion is unwritable under the old design.
The scheduler is the framing's state machine, all three review rounds
included: single-flight per group with generation-checked completions;
deadlines advanced from completion; the round-3 three-arm completion
partition (success / stale-or-retired / live non-success, with the
failure latch and quiet cancellation); joins wake the group, queue
exactly one follow-up mid-walk, and never reset the backoff curve;
per-watcher baselines --- the first snapshot whose WALK STARTED after
the join; membership captured at scan start; per-member cancellation
recheck at emit through the preserved _after_scan_for_tests seam;
backoff 250ms x2 to a 4s cap, reset by any emitted change; retirement
cancels the in-flight walk cooperatively.
Verification: eighteen acceptance tests. The six #234 tests are
byte-unchanged and green. Ten witnesses cover the framing's plan (the
review rounds added the fallback-determinism and root-boundary pair,
making twelve):
idle absence (and never a sleep purpose), one walk job per scan on a
twelve-directory fixture, join-wakes plus the registration epoch,
queued baseline for a mid-walk join (driven by saturating the worker
pool so the walk genuinely queues), single-flight under a withheld
completion pump, retirement and rebaseline through the fake's
unregister/re-register triggers, live cancel via pmacs._async._cancel
on the queued job, live failure with the once-per-error latch and the
preserved-snapshot recovery (DELETED for the pre-failure file is only
derivable from the retained snapshot), backoff shape from seam
timestamps, and the configured-root base.
Every witness was mutation-tested. Two findings from the bites:
- Retirement is DOUBLE-ENFORCED (unregister path and post-scan sweep)
and biting either copy alone is masked by the other; only biting
both goes red. Kept deliberately: the sweep covers seam-cancelled
members, the unregister path covers idle groups whose next deadline
is seconds away.
- The first idle probe was VACUOUS: it read pmacs.async instead of
pmacs._async, errored, and the unwrap_or_default made every sample
read as "absent". The probe now expects rather than defaults, so a
broken probe is a red test, not a green lie.
One environmental fact, recorded in the lane: an empty stray /tmp/.git
(since removed) made project detection root every markerless tempdir
fixture at /tmp, which under Q#D3-3 the watcher then faithfully
watched. A markerless-fixture red that looks like a watcher bug may be
an ancestor marker.
A pre-commit review round found four blockers, all fixed here:
- walk_tree checked cancellation only inside its entry loops, which an
EMPTY tree never enters --- a pre-cancelled queued walk returned an
empty SUCCESS, which the success arm would commit and diff into a
deletion storm. Cancellation is now checked before opening and
before returning, cancellation outranks a missing-root error, and a
unit test pins both.
- The neither-root-nor-cwd attachment fallback was still pairs-order
nondeterministic --- the exact accident D3 set out to remove, behind
a comment claiming otherwise. It now takes the lexicographically
smallest attachment directory. Verified at the spawn sites: every
server spawned with an attached file gets cwd = root, so the arm is
defensive and unreachable through production spawning --- which is
also why it carries no through-the-server witness.
- A base at the filesystem root joined as //path (and file:////path in
URIs). Both join sites now go through join_under, the root-aware
idiom dired's handler already uses, and the dir-of capture for a
root-level file ("" from the match) normalizes to "/".
- The walk-count and scan-times probes defaulted on error, so two
broken probes could compare equal and pass the retirement witness.
Every probe now expects --- a broken probe is a red test, the same
correction the vacuous idle probe forced.
A second pre-commit round found three more, all fixed here:
- Mid-walk cancellation was UNWITNESSED: both Rust cancel tests
pre-cancelled and the acceptance test cancelled a queued walk, so
deleting the internal polls left every test green. A cfg(test)
entry hook now flips the token at an exact entry boundary and the
witness asserts the walk stopped NEAR it (bound on entries
processed), which is what discriminates the polls from the
entry/exit checks. The retirement witness now holds a walk in
flight across the unregister and asserts the job settles cancelled.
- The "unreachable fallback" claim was WRONG: pmacs.lsp.spawn may
omit both cwd and root_uri, and ensure_server adopts such a live
server for markerless files (root_uri and key_uri both nil). The
lexicographic-minimum fallback now has a through-the-server
witness: five sibling directories, the minimum opened last ---
five, because with two the build's hash order coincided with the
lexicographic answer and the first-pairs bite survived.
- The root-boundary joins gained a witness through exported
production functions (the _deliver_status pattern): the matcher and
URI builder driven at base "/", where reverting either join_under
call makes the anchored glob refuse //hit and the URI grow a fourth
slash. No fixture can walk / for real.
Verification totals after both rounds: eight walk_tree unit tests,
eighteen acceptance tests (six byte-unchanged, twelve witnesses), all
mutation-verified.
No wire change, no PROTOCOL_VERSION bump; walk_tree is an fs binding.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Review round 2, findings P2a and P2b, plus P3's stale recovery
summaries. Three defects, and the fix is deliberately different in each
place because the constraint is.
## P2a — invalid UTF-8 bypassed the `purpose` diagnostic
`required_purpose` read the field with `value.to_str()?`. Lua strings
are BYTE strings, so `purpose = string.char(255)` is a value a caller
can write, and `?` surfaced mlua's generic conversion error BEFORE this
lane's own diagnostic was ever constructed: the caller was told neither
the field nor the rule.
**This is the third time this project has hit the class** — an unowned
Lua string converted with `?` ahead of the owned message; the
destination-capture lane corrected the same shape two rounds ago. It
refused before spawning and nothing leaked, so the defect was the
message, not the outcome. The conversion failure is now mapped onto
this function's own message, and the new acceptance row asserts on
message CONTENT so retyping the read as a bare `?` breaks the test
rather than silently degrading the error.
Auditing the rest of the lane's diff for the same class turned up
exactly one more: `_push_dispatch_name` took `name: String`, so a
registered handler name that was not valid UTF-8 failed at first
dispatch with mlua's generic message. It now takes `mlua::String` and
maps that failure onto an owned diagnostic naming the argument and the
rule. Those are the only two Lua-string reads this lane added; every
other binding it adds takes `()`.
## P2b — no safe display-text boundary. Two halves, two different fixes
### Handler names are refused at the source
`pmacs.workers.register` type-checked its name and nothing more, which
was defensible while the name died inside `dispatch`. It no longer dies
there: the ambient carries it into every job the handler allocates and
composes it into `purpose`, which `*workers*` and the modeline both
render. So it now gets `purpose`'s meaningful-value standard —
non-empty, not whitespace-only — plus control characters, which have no
legitimate place in a registered identifier.
### Purposes are ESCAPED at presentation, not rejected at the registry
A purpose may legitimately contain a newline: a filesystem path can, and
`pmacs-magit`'s spawn purpose is a whole argv. **This is the shape of
the `#228` decision, and it is consistent with it** — the one-line
constraint belongs to the surface that has it, not to the registry that
does not. There, `Command.description` stays free-form and the two
single-row consumers clip with `description_first_line`. Here the
equivalent is escaping rather than clipping, because a purpose's later
words are load-bearing: an argv's second word says which file, and a
clip would drop it silently.
`purpose_for_one_row` states the property it exists for: **a row must
not be able to forge another row.** It escapes `\n`, `\r`, `\t` and the
rest of the Unicode `Cc` class (which covers ESC, so a purpose cannot
open a terminal escape sequence either), borrows unchanged when there is
nothing to escape — making byte-identity structural rather than
asserted — and deliberately does NOT escape backslashes: no number of
them produces a second row, and doubling them would cost byte-identity
for ordinary text.
Two surfaces call it: the `*workers*` rows, and `ActivitySummary`, which
exists for one consumer that has exactly one row.
`pmacs.workers.snapshot()` is this lane's `describe-command` and stays
raw, which is what makes this a rendering decision rather than data
loss — asserted, not assumed.
## P3 — two stale recovery summaries
`docs/worker-identity-framing.md` still said "Implementation may
proceed"; it is implemented. `docs/active-work.md` still said Stage 1
takes the "first two" of owner/purpose/parent — `owner` was REMOVED in
revision 2, so it takes one of the three, and the claim the whole
`owner` argument overturned was still standing in the volatile state of
record. Both fixed section-locally.
## Verification
`tests/worker_identity_acceptance.rs`, 18 -> 24 tests:
* invalid-UTF-8 purpose refused by THIS lane's message, asserted on
content, alongside the absent / empty / whitespace / wrong-type /
metatable rows;
* a whitespace-only handler name and a control-character one are each
refused AT `register`, asserted on the error and on the handler not
being installed (dispatch reports `unknown handler`);
* a non-UTF-8 handler name is refused before the handler runs, with the
dispatch-name stack left empty;
* a purpose containing a newline renders as ONE row in `*workers*` and
as one line in the modeline — through the real rendering path, the
latter through a painted frame as well as the evaluator;
* **a purpose crafted to look like a row boundary does not produce a
second row** — asserted by counting rows, with the escaped text
asserted present so a renderer that dropped the purpose entirely could
not pass;
* a purpose with no control characters is byte-identical on both
surfaces, fixtured with a literal backslash, a literal `\v`, quotes
and a non-ASCII character.
Mutation-checked, seven guards, each failing its own test and no other:
the purpose UTF-8 diagnostic; the `_push_dispatch_name` one; the
register whitespace guard; the register control-character guard; the
`*workers*` call site; the `ActivitySummary` call site; and
`purpose_for_one_row` itself neutered to the identity, which fails both
surfaces' tests and nothing else.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016bqGA6s9tTUFzYpbeW3tai
`COHERENCE.md` §9 grades the worker model "mechanism without identity",
and §0 names step 11 (background-work ownership) as one of the two
remaining thin ends of the golden journey. The mechanism half is solid —
cancellation, supersession, streaming, frame-aware draining, `*workers*`.
The identity half was absent: `PendingJob` carried no description of what
it was doing, `pmacs.workers.dispatch` discarded the registered handler
name three layers above anything that takes one, and §9's "no progress
indicator exists anywhere" was checkable and true.
Framing: `docs/worker-identity-framing.md` (revision 4, approved).
What lands:
**A required `purpose`, on the job and on the process.** Non-optional,
with no `Default`, so the compiler — not a test — is what proves every
dispatcher supplied one. `allocate` / `allocate_with_resource` collapse
into ONE private `JobSpec`-taking funnel (Q#W-1): the two-function split
existed only because one prior lane needed one extra parameter, and a
second lane doing the same produces `allocate_with_resource_and_identity`.
`register_external` gains a `purpose` parameter rather than deriving one,
because its `JobKind` is `McpRequest`/`LspRequest` for every method — a
category, not a description.
**A dispatch-name ambient (Q#W-2), read at that same single funnel.** The
capture point is Rust, not the Lua wrapper layer, because a handler
reaching straight for `pmacs._async._dispatch_*` bypasses the wrappers
entirely — and those are precisely the callers attribution exists for.
Seven rules; the ones that decide whether it is honest:
- **Rule 1 — the extent is NON-YIELDABLE, and that is ENFORCED.** Both
supported yield APIs refuse inside it, modelled on the `commit_to`
refusal already in `async.lua`. The guards reject BEFORE parking and
reject UNCONDITIONALLY: one placed after `_is_complete` would fire only
when a yield really occurred, passing under test and failing
intermittently in production.
- **A raw `coroutine.yield` is NOT covered, and nothing here claims it
is.** R46 is a convention, and the scheduler inspects the yielded value
only after `coroutine.resume` returns — by which point the coroutine has
already suspended — so no refusal sited in a yield helper is ever
consulted. The residual is recorded in the framing §2 and in the
suite's module docs rather than papered over with a test that would
imply coverage this design lacks.
- **Rule 5 — unwind-safe.** A raising handler still pops. A version that
did not would let one failure poison every later dispatch in the session
with a stale name: the feature would stop failing loudly and start lying
silently. The bracketing also has to preserve the tail call it replaced:
`dispatch` was `return handler(args, opts)` and propagated EVERY return
value, so the pop/rethrow runs behind a varargs boundary rather than a
`local ok, result = pcall(...)` that would silently truncate a
multi-value handler. Varargs rather than `table.pack`, because that is
Lua 5.2 surface and LuaJIT is this project's default backend.
- **Rule 6 — compose, do not replace.** `"<name>: <purpose>"`, because
letting the dispatcher's purpose win loses the third party again and
letting the name win discards the only description of the actual work.
**A statusline activity indicator** — the fourth `pmacs.statusline.register`
adopter, after `mode`, `terminal` and `lsp`. A count plus the OLDEST
in-flight job's purpose ("busiest" is not a defined quantity; jobs carry
no cost estimate), and **absent entirely** when idle rather than a
zero-width segment that costs modeline width forever to say nothing is
happening. Gated by one setting, `ui.activity-indicator` (boolean, default
true, Q#W-6) — a permanently-visible modeline element is a preference
someone genuinely holds on day one. No setting for purpose capture
itself: that is substrate.
**NO WIRE CHANGE.** The indicator rides the existing `StatuslineSegments`
vector, so a fourth provider adds an element, not a variant.
`PROTOCOL_VERSION` and `ADVERTISED_PROTOCOL_VERSION` are untouched — which
is the property that lets this run beside the two lanes holding the bump
slot.
**Q#W-7 — a pre-existing defect, repaired here, and NOT one anybody has
observed.** `Handle:await()` refuses inside `pmacs.window.commit_to`
precisely so a coroutine cannot park with the frontend scope pushed
(Journey Stage 1a, Q#JR14b). But `pmacs.async.yield_to_next_tick()` also
yields, is public, and carried no such refusal — so that invariant had a
second entrance, and a coroutine could produce exactly the misrouting the
`await` guard exists to prevent. It gains both refusals here: the same
supported yield helper, the same invariant, the same edit family, so
splitting it would have preserved a known hole without reducing
integration risk.
**Reachability by a real caller is UNPROVEN.** This was found by reading
the guard family while scouting rule 1, not by reproducing a fault. No
production caller is known to yield through that door inside a commit,
and the test pins the guard rather than reproducing a user-visible bug.
Nobody should later cite this commit as evidence the bug was observed in
the wild. Its witness is a PAIR, like rule 1's: the refusal fires **and**
the commit scope is restored afterwards — a guard that raises while
leaving the scope pushed converts a silent fault into a loud one and
fixes neither.
`journey_acceptance` carries the established `commit_to` pins —
forged-destination refusal, scope-and-restore on normal return and on
raise, the await refusal, delivery to the requesting frontend. It passes
**untouched**, which is what says this closed a gap in Journey Stage 1a's
semantics rather than altering them.
What is deliberately NOT here, and why it is worth saying:
- **No `owner`, in any spelling** — not `origin`, not `subsystem` (§3).
Populated from static per-subsystem constants it would be an origin,
not an owner, and would confidently misattribute third-party work to a
builtin at exactly the point §9 wants attribution. A field that asserts
a falsehood is worse than an absent one. The slot stays empty until P3
can fill it with a real package signal.
- **No `parent`** (Q#W-5). An unpopulated field renders as `None`
everywhere and reads as "this job has no parent" rather than "this
system does not track parents". Stage 3 builds the lifetime model and
the field together.
Consequences worth recording:
- `ProcessSpec::new` takes a third argument. The 40-odd call sites are
almost all tests; the three production ones (LSP, MCP, terminal) supply
real descriptions. `pmacs.process.spawn`'s Lua surface keeps `purpose`
OPTIONAL, falling back to the label — requiring it there would break
every existing caller for no coverage the compiler is not already
providing, and a caller's own label is not a fabrication.
- `pmacs.process.list` gains a `purpose` KEY on each row and enumerates
exactly the same processes (Q#W-4). Terminal PTYs stay hidden: three
acceptance suites use `#pmacs.process.list()` as a leak baseline, and
widening the accessor would inflate all three. Stage 2's unified view
owns that decision.
- `statusline_segments_acceptance`'s builtin-provider inventory grows to
`["activity", "mode", "terminal", "lsp"]`. That assertion exists to
grow when a builtin provider is added.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016bqGA6s9tTUFzYpbeW3tai
`src/lsp.rs` gains nine: the fourteen-family store inventory with a
17-entry precondition so it cannot pass vacuously, the route purge with
`workspace/symbol` and another server's route both surviving, the
awaiter drain joined on the rid, the error contract's two arms, the
late-publish drop with its does-not-over-reach companion, the
`mark_document_stale` gate across all three stale stores, exact-pair
tombstone identity, and reclamation under both `start_generation` and
terminal `forget`.
`src/diag.rs` pins that `forget` drops the epoch while `clear`
deliberately bumps it — the leak a `clear`-based forget would leave in
the one map nothing prunes.
`src/async_runtime.rs` injects two resource replies onto the private bus
in each order and asserts `TickOutcome.resources` reports arrival order,
not allocation order; plus that a failed or cancelled mutation is not
harvested at all.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Lv428Fth9LRtffwJSsqH7T
One shared walk query (`buffers_bound_under`), lifted out of #190's
`delete_verdict` so the guard and both reconciliation seams cannot
disagree about which buffers an operation touches: every buffer, both
sides normalized, component-aware containment.
`EditorCore::reconcile_rename` moves the stored path and — only for a
`PathDerived` name — the buffer name. `EditorCore::reconcile_delete`
composes the same two removal phases `pmacs.buffer.kill` composes,
preflighting `editing_in_progress` because a `ConcurrentEdit` refusal
arrives after `kill_buffer` has already moved windows. Phase 2 stays
with the caller; `EditorCore` gains no Lua handle.
`AsyncRuntime::tick` now returns a `TickOutcome` carrying the settled
ids plus the successful resource mutations, in bus-arrival order, which
is documented as not being execution order. `PendingJob.resource`
retains the paths the dispatchers move into the worker closure.
`LspManager::forget_uri` purges the routes carrying a URI, drains the
awaiters joined to them on the rid, and clears all fourteen stores plus
`documents`. A generation-scoped exact-pair tombstone gates the two
uncorrelated writers that can otherwise resurrect what it cleared:
`publishDiagnostics` and `mark_document_stale`, which now takes a
server id. `ResponseRoute::scoped_uri` is the one variant list, with
`uri()` delegating to it.
New Lua surface: `pmacs.buffer.set_name`, `pmacs.lsp.forget_uri`,
`pmacs.diag._rename_resource`.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Lv428Fth9LRtffwJSsqH7T
Dired is the file surface, not a rider on one: before Stage 0 (#162)
pmacs had no way to open a file by path, and browsing is the half a
user reaches for when they do not already know the path. Stage 1 ships
the view.
builtin/runtime/dired.lua: one buffer per directory named by the
canonical path (Q#DR2) with an ownership check before any paint (F7);
read-only intercept plus round-trip input (Q#DR3); a `dired` major mode
carrying mode-scoped keys (Q#DR8) -- RET/f visit, ^ parent, n/p, g
revert, q quit, s sort; cursor re-seated by basename across every
wholesale repaint (Q#DR9); file visits through
`pmacs.window.display_file` and directory descent through dired's own
window (Q#DR10); `C-x d` / `C-x C-j`; and `dired.kill-when-opening`
through the config registry.
Two Rust changes, both narrow:
* `read_dir` grows per-entry tolerance behind an opt (Q#DR6). Five
per-entry conditions used to fail the entire listing, so a plain
refresh of a busy directory could just fail; the module doc's claim
that a tolerant wrapper was "the package's job" was false, because
the primitive hands Lua one structured error and no partial vec.
Per-entry readdir/lstat/readlink failures and non-UTF-8 symlink
targets now land in an `errors` channel; parent-level failures and
non-UTF-8 *names* stay fatal. The tolerance travels in the settled
payload, so the Lua boundary keeps the bare-array shape the frozen
M8.2 fixture consumes and never has to look the job back up. The read
ops' opts parsing now rejects unknown keys, so a typo'd `tolerant`
cannot silently degrade to the fatal contract.
* `normalize_buffer_path` is exposed as `pmacs.path.canonicalize`
rather than mirrored in Lua. Q#DR2 named exposure the preferred end
state; it needs no borrow plumbing, so dired's name-dedup and
`display_file`'s `find_buffer_for_path` dedup cannot fork, and the
mirror's Stage 2 removal is not owed.
tests/dired_acceptance.rs covers framing items 1-16 (22 tests), driven
through real key dispatch. Item 17 is the m8_1/m8_2/m8_3 gate.
One framing claim is corrected by the substrate: R2-3 expected a
dedicated dired panel to carry its dedication across a descent, but
`display_buffer` never replaces the buffer in a slot dedicated to
another one -- it discards every side-specific parameter and falls back
to the document window (Q#BP3 2.iii). Dired does not try to unpin the
user's panel; both arms are pinned.
Teach the syntax engine that one buffer can hold more than one
language. After the root parse, run the grammar's injections.scm, parse
each embedded region with the injected language, and merge every
layer's highlight spans. First consumer: markdown fenced code + inline
(zero new grammars — the block grammar already ships an injection query
and the injected langs already have grammars from #118).
Engine (src/syntax.rs):
- ParseTreeBundle now holds Vec<Layer> (root layer 0 + injected
children, depth-ascending); installed atomically so the existing
Arc::ptr_eq style gate and highlight cache keep working (Q#IJ1).
- run_parse builds layers on the worker: run injections.scm, resolve
the injected language, compute Vec<Range> (exclude NAMED children,
intersect the parent's ranges), set_included_ranges cold-parse,
recurse — bounded by depth (3), a layer backstop (4096), and a
(lang,ranges) visited guard; any child failure drops that child only
(Q#IJ3/IJ5). LanguageEntry gains injections_query; markdown_inline is
registered (retires the M9.7 block-only floor); markdown/rust carry
injection queries.
- Injected languages resolve off the static BUILTIN_LANGUAGES table
(Send loaders + query sources), preserving lazy loading. Dynamic
fence names go through a case-folded alias map seeded with defaults
and Lua-extensible via pmacs.parse.injection_aliases, snapshotted into
ParseRequest at dispatch so the worker never touches the Rc registry
or a Lua table (Q#IJ2/IJ4). Highlight queries are resolved at settle
(resolve_layer_queries), keeping query compilation main-thread/cached.
Producers:
- SyntaxHighlightView (grid) iterates layers shallow-to-deep so a
deeper layer's styling wins within its region (Q#IJ6/IJ7).
- scoped_style_spans (wire) flattens all layers into DISJOINT effective
spans via a boundary sweep, since the GPU re-sorts spans by start
(replace_style_spans / merge_style_spans) and would otherwise destroy
producer order. The GPU source_color_at consumer is fixed to fold all
covering spans (matching semantic_client's effective_style_at) rather
than returning the first.
Named-children exclusion: content ranges exclude only NAMED children
(matching tree-sitter-md's own inline splitter) — excluding a block
inline node's anonymous text tokens would shred the paragraph into
unparseable fragments.
13 acceptance gates (framing docs/multi-language-injections-framing.md):
layer structure, absolute child offsets, alias resolution (static +
case-folded dynamic + unknown-skip + Lua-async override), multi-range
inline, recursion bounds, wire + grid + GPU producers, incremental edit
/ new fence, many-paragraph settle budget with tail coverage, and the
single-layer regression guard.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YJ9FQ832QwftJXCD9LeFan
Every EditorState leaked its entire worker pool (cores-1 threads, each
waking every 100ms): the Rc<AsyncRuntime> is cloned into dozens of Lua
closures, and registries those closures capture store mlua::Function
values -- reference cycles through the Lua VM that keep the Rc from
ever reaching zero, so WorkerPool::Drop never ran. Harmless for one
editor per process; in the m4 acceptance suite (54 editor-building
tests) it accumulated 1000+ live threads (observed: 60 complete
pmacs-worker-0..14 pools at once) plus ~9k spurious wakeups/second.
Fix: WorkerPool::signal_shutdown() -- set the shutdown flag + wake the
parkers from a shared reference; parked workers exit within their
100ms park timeout. Exposed as AsyncRuntime::shutdown_workers(), called
from a new impl Drop for EditorState. Deliberately signal-only, NO
join on the drop path: a worker can be blocked publishing its reply
onto the bus only the main thread drains, and a first cut that joined
in Drop deadlocked the m4 suite at teardown (fake-LSP tests wedged
2h+). WorkerPool::shutdown() (signal + join) remains for owners with
no bus consumer to deadlock against; Drop delegates to it as before.
Measured on the m4 suite: peak live threads 1026 -> 122.
Regression: tests/worker_shutdown_acceptance.rs (thread-count probe is
/proc-based, Linux-gated; the idempotence test runs everywhere).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Replaces the editor-blocking `poll_until` tick-loop in the LSP UX
with the M9.1 external-settle pattern: each `textDocument/*` request
registers a pending entry via `AsyncRuntime::register_external` and
returns the job id; the JSON-RPC response (or a server-teardown /
cancel / timeout) settles it, resuming a `Handle:await()` coroutine.
No worker thread is occupied for the round-trip.
Hybrid result delivery (operator decision): the response is absorbed
into the typed stores *and* carried through the Handle. The
completion popup and diagnostics gutter keep reading the stores
untouched; request/response command code awaits the value directly.
Core (src/lsp.rs):
- `LspManager` gains `runtime: SharedAsyncRuntime` (threaded through
`make_lsp_manager` / editor.rs, mirroring `make_mcp_manager`) plus
a `(server, request_id)` -> PendingExternal awaiter table parallel
to `pending_routes`.
- `request_*` return the async `JobId` (`= u64`, signature
unchanged; no caller consumed the old JSON-RPC id).
- `handle_response` settles every non-cancelled awaiter ok/failed
alongside store absorption; null result still wakes await with nil.
- Awaiters drain-cancelled at all three `pending_routes` purge sites
(restart generation flip / terminal exit / forget) so a coroutine
cannot park on a server that went away.
- Per-tick sweep: per-awaiter cancellation (Handle:cancel() or
supersede via a stable `lsp:{method}:{sid}:{uri}` key), with
`$/cancelRequest` + `cancelled_rids` on abandonment to drop the
cancel/response race silently. Mirrors mcp.rs.
- Per-request timeout (default 10s, `pmacs.lsp.set_request_timeout_ms`):
an alive-but-silent server fails the await instead of hanging.
Lua surface:
- `_request_*_raw` job-id bindings (mirror `pmacs.mcp._send_request_raw`).
- builtin/runtime/lsp.lua: Handle wrappers + the four commands
rewritten to spawn `pmacs.async` coroutines that `:await()`;
`poll_until` removed. Server-gone / error surface as structured
await failures in the modeline.
Tests:
- pmacs_fake_lsp: `error` / `silent` modes for deterministic
failure-path coverage.
- 5 end-to-end await-path tests (success+store, server-error->failed,
server-stop->cancelled, timeout->failed, supersede->cancelled).
Gate: fmt clean; clippy --all-targets -D warnings clean; lib
1223/0; m4_acceptance 58/0; m9_1_acceptance 18/0 (MCP unaffected).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Root cause of the CI Lint regression: commit 6113c53 bumped
rust-version 1.85 -> 1.95. clippy::collapsible_if is MSRV-gated —
collapsing `if let { if let }` needs let-chains, stabilized in Rust
1.95. At MSRV 1.85 clippy suppressed these; at 1.95 it emits them.
The patterns were pre-existing; the MSRV bump surfaced 47 of them
and turned `Lint (luajit)` / `Lint (lua54)` red at HEAD (was green
through PR #7; red from PR #8 = the release-prep MSRV bump).
Resolution (operator-chosen: autofix into let-chains): applied
`cargo clippy --fix` across the luajit, lua54, and crdt lanes
(--all-targets). The fix only applied with the lint at warn level;
`-- -D warnings` turns it into an error and blocks --fix.
Verified on the pinned 1.95.0, all three lanes:
clippy --all-targets -D warnings clean (luajit / lua54 / crdt);
fmt 0 diffs; lib tests 1223/0.
Note: the prior #6 "quiescent audit, clippy clean" was inaccurate —
clippy was not actually re-run there (build/version/fmt only), so
this MSRV-gated regression went uncaught until the live attach-debug
investigation surfaced it. This commit restores genuine clippy
cleanliness at MSRV 1.95.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Land the Model Context Protocol (MCP) integration as a transport binding,
not a built-in feature. Six Lua functions plus userdata methods expose
the substance of three MCP feature areas (resources, tools, prompts), a
notification dispatcher, and a non-trivial AI-assistance example package
that meets the architectural ship gate (spec/pmacs-spec.tex:1572): zero
direct calls into the Rust core, zero special-cased MCP handling outside
the public API, source under 2000 lines of Lua.
The M9.5 -> M9.6 -> M9.7 -> M9.8 layered composition validates the claim
"AI is a transport binding, not a feature" -- pmacs-mcp-ai composes with
pmacs-mcp-prompts.render and inherits notification handling transitively
through M9.7's package, demonstrating that the AI domain is a layer
above MCP, not a thread woven through the core.
Subtask shape:
M9.1 stdio transport + initialize handshake + restart policy
M9.2 resources with in-flight + settled cache and per-uri invalidation
M9.3 tools with isError-vs-JSON-RPC-error semantics + cancellation
M9.4 prompts with required-argument validation
M9.5 notification dispatcher (on_notification, off_notification)
M9.6 tools-as-commands fixture package + 12 audit findings disposed
M9.7 prompts-as-result-buffers fixture package + tree-sitter-md grammar
M9.8 AI-assistance fixture package (363+ LoC; 17/17 acceptance tests)
M9.9 formal package audit -- PASS on all three criteria
M9.10 release: TRANSITION-M9.md + MCP-for-package-authors guide
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Levi Neuwirth