Doc 20 §6 D1, first slice: wasm/sched/context.{h,cpp} implements
create / yield_park / mark_ready / drain with the registry as truth.
The shape that matters — a STAR, not libcontext's symmetric swap:
contexts only ever swap OUT to the scheduler, and only the scheduler
swaps IN. So "is this target safe to enter?" stops being a guess
(libcontext's swap_suspended / parked / hot-main refusals) and becomes a
lookup: the registry says Parked/Ready and holds the buffer. Doc 19's
refused-resume is unrepresentable here because resume is not a decision
made at the swap site.
Enforced, not hoped for:
- at most one transition in flight (drain() refuses re-entry, so a
context calling drain() cannot turn the star into a cycle);
- mark_ready() never resumes inline — it queues, and drain() resumes
from a clean stack (doc 13 §1.4's deferred-wake law, per context);
- yield_park() off a context is REFUSED, which is the "nothing parks in
place" rule made mechanical;
- destroy() on a non-Finished context is refused (freeing a parked
stack strands whatever is on it);
- FIFO ready queue (no starvation);
- main-thread-only, per doc 21 §2's pthread finding.
Memory is accounted from the start (doc 20 risk 1): live/peak contexts,
bytes/peak bytes, and per-context asyncify high-water measured from
asyncify_data.stack_ptr — the same quantity the shim reports as `rem=`,
from the other end. Sizes deliberately start at 128 KB C stack + 128 KB
asyncify buffer rather than inheriting libcontext's 512 K, so the number
gets derived from evidence; a >75% buffer use beacons BUFFER-PRESSURE
because that overflow is silent corruption, not a crash.
No production path runs on this yet — dispatch moves at D2.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TEHGiiXMShNXbBr7gSJ7iz
Doc 19: status DIAGNOSED -> DIAGNOSED + PINNED RED; the repro section
leads with the automated deterministic spec (the manual Leonardo flow
stays as the field repro).
Doc 20: D0 work-log entry; status D-1 + D0 done; next = D1.
Doc 21: red-spec section updated to "landed", with the staging test's
anti-vacuity role.
Carried finding: the strand reproduces on a 2-OBJECT fixture schematic,
so warm-load byte volume (the 68/1 dice-loader) is not an ingredient —
two concurrent parks suffice. That is why D3 closes this class and D5 is
not required for it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TEHGiiXMShNXbBr7gSJ7iz
Doc 20 D0 second deliverable (doc 21 §4). tests/kicad/quasimodal-strand.spec.ts
reproduces the Symbol Properties hang on demand, in two parts:
- "staging" (GREEN): double-click the fixture symbol → Symbol Properties
opens → arm the parking timer (wasm/bindings/timer_park.h) WHILE the
dialog is up. Because the opener's fiber park is open-ended, a timer
firing now necessarily parks on top of it — the overlap is structural,
not a won race. Asserts the dialog opened, the timer fired, its OK
button is hittable, and the shim beaconed concurrent contexts. Keeps
the red pin from rotting into vacuity, and fails loudly on its own.
- "doc-19 red" (test.fail()): clicks OK and asserts the desired end state
— dialog closes, zero fiber-resume-refused beacons, wait books balanced
(no unresolved nested/modal wait). Goes green at D3, when Playwright
will report "expected to fail but passed" and the marker comes off.
Verified 6/6 consecutive full-file runs, identical outcome each time:
closed=false dialogs=1 refused-resumes=1 — the doc-19 mechanism exactly
(quarantined fiber's legitimate resume refused, dialog never closes).
Deliberately NOT asserted in staging: the timer park COMPLETING (whether
a park survives the aliasing is the disease under test) and sawParked (a
100ms sampler can miss a short park) — both are reported, not gated.
Retires tests/kicad/dialog-deadlock-probe.spec.ts: the 8/4 throwaway probe
that established the mechanism. This spec supersedes it and, unlike it, is
deterministic (the probe's 3 blind waitForTimeouts were the only
determinism-lint violations in the tree; the guard is now clean).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TEHGiiXMShNXbBr7gSJ7iz
Doc 20 D0 first deliverable: every Asyncify park site post-D-1,
classified by whose stack it suspends (tool fiber / entry stack / main
loop) with a routing decision per site — 14 production sites (8 wx, 9
KiCad/bridge counting quartets/pairs) + 3 deliberate test levers.
W1 (wxWasmYieldUntilJs) and W3 (popup) route to D3 context yields;
the clipboard/font/lib/3D/occ/ngspice/nanosleep bridges route to D4;
W2 (per-frame yield) stays safe-by-construction unless D5 is taken.
Also settles doc 20 risk 4 (pthreads): all parks are main-thread only;
the lib bridge's pthread path is blocking-proxy, not Asyncify, and is
out of migration scope. Defines what D4's "no handleSleep on a fiber
stack" assertion must cover.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TEHGiiXMShNXbBr7gSJ7iz
Gate results (single battery, scheduler-only): wx app battery +
asyncify + coroutine 363 passed / 3 skipped / 0 failed; full kicad
suite 138 passed / 30 skipped / 1 failed — the one failure is the
pre-existing local occ-probe glb case (predates D-1, unrelated).
Post-gate scrubs: vestigial 'startModal' removed from ASYNCIFY_IMPORTS
(tests/apps/Makefile.wasm) and env.startModal from
scripts/common/asyncify-imports.txt (the import no longer exists in any
wasm; both lists are boundary supersets so behavior is identical).
Stale comment pointers to the deleted legacy modal machinery updated.
Doc 20: D-1 work log added, status flipped to IN PROGRESS.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TEHGiiXMShNXbBr7gSJ7iz
Doc 20 D-1, second slice: the injector now injects asyncify-scheduler.js
unconditionally — the WX_SCHEDULER=0 legacy opt-out, the
SHIM_DISABLE_HANDLESLEEP ablation skip, and the
SHIM_DISABLE_TRAMPOLINE_HEAL ablation skip are gone, and
scripts/common/shims/handlesleep.js is deleted (the scheduler subsumed
its capture/restore, fiber guard, and trampoline-heal duties in S2).
Comment-only scrubs point the remaining references at the scheduler
shim. .ci-cache-epoch bumped (shim/injector behavior changed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TEHGiiXMShNXbBr7gSJ7iz
Doc 20 D-1 (legacy runtime deletion), first slice: races_test_noheal /
races_test_nosleepfix pinned behavior of the legacy handlesleep shim,
which is being deleted — the pins now assert properties of a runtime
that no longer exists. Drops the two Makefile.wasm link+inject variants
(SHIM_DISABLE_TRAMPOLINE_HEAL / SHIM_DISABLE_HANDLESLEEP), the
shim-redundancy pin specs in asyncify-races.spec.ts, and resolves the
tests/README.md open task. The green battery still runs every scenario
against the scheduler glue.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TEHGiiXMShNXbBr7gSJ7iz
Carrying the WX_SCHEDULER=0 runtime through the core rewrite would make
every phase dual-path (two code paths per park site, dual-glue builds,
two batteries per gate) for a runtime we intend to delete anyway; git
on the feature branch already provides the rollback the fallback was
for. D-1 now deletes it up front — injector branch, handlesleep.js,
the wxWasmMailboxEnabled gating, the wx legacy twins, and the ablation
builds that pin the old shim — executing doc 17's S5 ledger item 1
early. Interlock and busy gates unaffected (ledger items 2-3, D2/D6).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TfxKn5utcntBSnxz4ZnYKs
Doc 20: the remaining core of Design B, scoped against what S0-S6 built
and motivated by the doc-19 hang. Core rule: nothing parks in place —
every suspendable activity yields its own context, so a context's state
is authoritative and the guessing layer (quarantine, consume-once,
libcontext refusals, dispatch interlock) gets deleted rather than
tuned. Phases D0-D6 with gates, 6-10 wk; memory/partial-migration/
pthread risks called out; index updated.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TfxKn5utcntBSnxz4ZnYKs
Reproduced live on the dev platform: the tool fiber running the
quasi-modal parks mid-body, is quarantined by the stale-fiber guard,
and its resume is REFUSED — so it never releases the dispatch guard.
Interlock held forever => clicks deferred and never drained, timer
delivery frozen; the titlebar X works because it is ungated.
Includes the captured frozen state, what is ruled out (clicks do reach
wx; no I/O in flight), and ranked fix directions. Regression vs
pre-existing still undetermined — needs a real WX_SCHEDULER=0 build.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TfxKn5utcntBSnxz4ZnYKs
Injector defaults to asyncify-scheduler.js (WX_SCHEDULER=0 = explicit
legacy opt-out); .ci-cache-epoch 9->10. Doc 17 S5 corrected: the
interlock/busy-gate deletions assumed handler-fibers that S1-S4 never
built — they stay as load-bearing second lines; each real deletion is
ledgered with its unlock condition. Flip gate: full kicad suite 136
passed on BOTH variants (occ-probe glb fails identically on both =
pre-existing; ngspice bg_run = rerun-passes flake).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TfxKn5utcntBSnxz4ZnYKs
asyncify-scheduler.js REPLACES handlesleep.js on WX_SCHEDULER=1 builds
(injector either-or): ports capture/restore, fiber consume-once/
quarantine guard, wake-window flags, recorder, trampoline heal — and
adds deferred wakes (a wake mid-transition queues and drains from a
clean macrotask) plus the N1 currData accessor (pure-JS writes need
scheduler authorization; strict mode throws; meta-tested). Gates:
races 9/9 with NO legacy shim (subsumption), coroutine 39/39,
wx-chromium 30/30, kicad trio 3/3 on the C-lane build.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TfxKn5utcntBSnxz4ZnYKs
Shim embind lane wraps the doc-18 production mutators at the Module
boundary (busy-window calls queue + deliver post-settle; time-boxed
unkillable pump). N2 un-fixme'd and green; collab-load-fuzz carries the
variant contract (drop on legacy, deliver-in-order on scheduler, capped
hammer on the scheduler lane); timer-park's timerRetry silence tripwire
arms on shim+export and is green on the C-lane kicad build. Bump
wxwidgets for the wheel lane. CI both-EH matrix deliberately deferred.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TfxKn5utcntBSnxz4ZnYKs
Doc 17 step S0 scaffolding: WX_SCHEDULER=1 injector path with an
observation-only asyncify-scheduler.js skeleton (legacy shim stays
authoritative until S2), guard-beacon extraction with occurrence
recovery for rate-limited beacons, and the fixme'd N2 ordering spec
(add-then-move probe; un-fixme at S1).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TfxKn5utcntBSnxz4ZnYKs
Doc 17: Design B phasing revised with the July-August guard record.
Test inventory with per-test fate (keep / rewrite / retire / new),
steps S0-S6 with gates and rollback, ~5-7 wk.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TfxKn5utcntBSnxz4ZnYKs
Picks up the wxwidgets correction for the CI failure: the loop-depth gate is
replaced by releasing the parked nested DoRun from the scheduled tick's own
error path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
Picks up the wxwidgets follow-up CI caught: the top-level loop must not
schedule dispatches while a quasi-modal's nested pump owns event delivery.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
The editor reported nothing when a session died. Evidence lived only in-tab —
an 800-line React array behind a "Show console" button — so diagnosis meant
asking a user to paste a screenshot.
Better Stack's Error Tracking ingests the Sentry wire protocol, so this runs
the stock @sentry/browser against a Better Stack DSN. Sentry.init installs its
own window error/unhandledrejection handlers, so uncaught main-thread errors
and the wasm traps that escape emscripten's DOM event handlers are captured
with no instrumentation at the throw sites. Not their JS tag: it has no
beforeSend or fingerprint hooks, its runtime spawns workers from cross-origin
CDN hosts (this page is COEP: require-corp), and it ships session replay on by
default — which on a CAD canvas records customers' board geometry.
@sentry/browser is imported in exactly one file so the vendor stays swappable,
mirroring how lib/analytics.ts isolates Plausible.
Also replaces the terminal-signature regex with a shared, unit-tested predicate
(wasm/terminal-error.ts) used by BOTH the fatal overlay and the reporter, so
they cannot disagree. The regex was a type check written as a string match and
had three live holes: `RuntimeError` was listed but never appears IN
`.message`; Chrome's bare "unreachable" and "null function" matched nothing
(the v0.1.20 prod log is exactly those); and narrowing "table index is out of
bounds" to `\bindex out of bounds` for Firefox in 197f317 silently stopped
matching Chrome's spelling. Checking the TYPE — every trap in this family is a
WebAssembly.RuntimeError — covers all engines and ends the spelling chase; the
message patterns remain as a fallback for paths that lose the Error object,
such as a worker ErrorEvent crossing the realm boundary with error: null.
197f317's pthread-worker tap, promote() and Firefox findings are kept as-is.
Notes:
- Off unless VITE_ERRORS_DSN is set AND VITE_ALLOW_USER_OVERRIDE !== "1" (dev
servers and every Playwright harness set the latter, and production builds
never do), so a production DSN in a local .env still cannot report. With no
DSN the whole SDK is const-folded out: 1,193,080 vs 1,282,463 bytes of JS.
- browserApiErrors integration removed. It wraps setTimeout/rAF/addEventListener
in try/catch, which is exactly how KiCad-on-Emscripten drives its main loop.
- Console breadcrumbs off (collab/debug.ts's clog fires per Yjs update and would
evict the ring before any crash); dom/fetch/navigation breadcrumbs kept.
- beforeSend redacts token/apiKey/Bearer — collab/provider.ts puts the collab
token in the y-partyserver URL, so a connection-failure string carries a live
credential — and guards the cascade: one wedge produced 8 errors in prod, and
after the first terminal event the rest are dropped into cascade_count.
Verified end to end against the real EU host from a cross-origin-isolated page:
POST /api/<id>/envelope/ -> 200, and 4 terminal throws produce 1 event
(control: 1 throw, same count).
Privacy policy 9, cookie policy 6 and the licenses page are updated: Better
Stack is disclosed as an EU processor, and the licenses page now describes the
browser app's own JS dependencies, which it never did.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every repo moved from emergence-engineering/ to PCBJam/, but the published
source pointers still named the old org. GitHub's transfer redirect resolves
them (all six checked, 301 -> 200), so nothing is broken today — but it stops
working the moment anyone creates a repo at an old path, and a GPLv3
corresponding-source pointer is a poor thing to leave depending on a redirect.
CI never passes --repo, so the hardcoded default is what actually ships: the
editor's version badge on editor.pcbjam.com has been linking users to the old
org for their source.
Covers the source pointers (licenses.md, terms.md 12.4, REPO_URL and its doc
comment, the three build-script --repo defaults, the site footer's build-commit
link) and the two "our GitHub" org links. Bumps the pcbjam-shared pointer for
the same fix there.
Deliberately untouched: emergence-engineering.com, the company domain behind
contact@ and the EE credit block — the trailing slash in the substitution keeps
it out. And docs/security-audit-glm/, which describes a finding rather than
linking anywhere.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Picks up the wxwidgets fix (main loop schedules its events into a fresh JS
task instead of dispatching them inside its own Asyncify wake continuation)
and documents the whole round in docs/features/async/16.
Local verification on the warm-load repro built yesterday — the case that
failed every warm load on every build since v0.1.12: 3/3 loads settle with a
fully rendered Leonardo board, rootHotTotal=0 (it was exactly 1 at every
death), fcsTotal=72, no traps. Full kicad e2e: 136 passed, 1 failed, and that
one (occ-probe's GLB format) fails identically on a build without this change
— a pre-existing OCC build-flag issue, tracked separately.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
The prod differential ladder finished: staged byte VOLUME on a warm load is
the only trigger left (V1a siblings-without-lib-tables dies, V1b +120 files
survives, V1c sibling KiCad files renamed byte-for-byte dies, V1d Leonardo +
123MB of inert markdown dies on loads 3-4; 14MB never dies). 3D models,
collab/ydoc/presence, lib tables, sibling KiCad handling and file count are
all exonerated — volume only loads the dice on the underlying race.
That made the crash reproducible locally for the first time in six campaigns:
a persistent browser profile + a 110MB project fails every warm load with the
exact prod signature. Iteration is now ~12 minutes instead of a release cycle.
Shim: every fiber switch now records the departing side's remaining asyncify
buffer and its recorded rewind entry (rem=/rf=), which is what identified the
unrewindable capture and disproved buffer overflow. The deferral family is
closed for good — a microtask-deferred retry on a clean empty stack died
identically to the nested rewind, because the suspension is broken at write
time, not by nesting.
Shell: log the origin stack when wx reports the top window destroyed. That
notification fires from ~wxTopLevelWindowWasm for ANY top-level window, so a
transient frame dying mid-load navigates the user out of the editor — a real
bug in its own right, found while chasing the empty flight-recorder dumps.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
The 8/2 crash-hunt bisection attempt with ?collab=0 was silently invalid: the
flag only gated the attach, while the boot fan-out joined the doc room and
materialized the target file from the ydoc regardless (the "collab=0" prod
log shows both, plus an attach). The flag now also skips the doc-room join —
the file falls back to the plain fetch path — making it a real lever for the
warm-siblings crash bisection (ydoc-vs-sexpr file source, the next suspect
after sibling restage and collab attach were exonerated) and an honest
user-facing escape hatch.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
The differential-repro ladder (2026-08-02, five prod runs + local counter
measurements) narrowed the crash trigger empirically: warm loads of
sibling-heavy projects die at settle (V1/V4 fail 2/2 warm; V2/V3 without
siblings never fail, warm or cold), while the flight-recorder counters show
the settle-time collision windows themselves are universal (fcsTotal=72,
rootHotTotal=3 on V1 AND V3, every load, cold and warm — so the windows are
the shared fan-out, not sibling-made). The sibling restage's room connects +
restage fetches are the only sibling-specific traffic contending with those
windows, and warm IDB compresses it into exactly that moment.
Nothing in the restage is needed for first paint — the boot snapshot staged
every sibling seconds earlier — so it now starts on requestIdleCallback
(5s timeout; setTimeout(3s) fallback), well clear of the settle storm.
Unmount-safe via disposedRef (armed per mount, checked in the deferred
starter and on handle resolution).
Validation is empirical by design: the counters won't move (windows are not
sibling-made); the test is warm V1/V4 prod loads no longer dying. If they
still die, the sibling lever is exonerated too and the remaining suspects
narrow to the ydoc-materialization path差 (second-load file source) — the
next probe either way.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
The 96-event recorder ring holds under a second of history at idle tick rate
(~110Hz), so settle-time collision pressure scrolls out before any poll can
read it — the prod dumps only caught the kill because the trap froze the
moment. Scroll-proof totals since boot: every finishContextSwitch increments
fcsTotal; every root entry inside a sleep-wake window increments rootHotTotal
(the fatal precondition). Both appear in the [wx-asyncify] STATE line, i.e.
in every trap auto-dump, every __wxAsyncifyDump() call, and every blue-screen
console — turning ANY prod load (crashing or clean) into a dose measurement
for the differential-repro experiment (which project ingredient generates
collision windows: siblings, 3D models, libs).
.ci-cache-epoch 7→8.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
Crash-hunt sessions kept re-asking "was that even the new build?". Every boot
now logs two self-identifying lines into the in-app log (and thus the fatal
ring + blue screens): the app's own bundled chunk name + manifest base + UA,
and the wasm's CDN ETag + SHA-256 of the first 128KiB + length — teed from
the same stream the progress counter reads (no second download, no
buffering), emitted early so the line exists even when the load dies later.
Verify against the CDN with: curl -r 0-131071 <wasm-url> | shasum -a 256.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ac/code/session_019SE4o46Lnq3hF574FFq8x4
v0.1.24 in prod, doubly convicted the same morning: the Leonardo open
crawled/hung (open:settled result=failed at the 60s escape, heap never past
256MB — every main-loop iteration runs INSIDE its yield-wake extent, so the
"root-owned wake" scope matched thousands of legitimate nested coroutine
Call/returns per open, each paying a deferred macrotask, throttled to ≥1s in
a background tab), AND the Nano crashed 22ms after deferrals=1 fired.
Harmful and insufficient: the fatal nested-rewind interleave and the benign
bulk are observationally identical at this layer — no discriminator exists.
Retired (second and final retraction, async/16 round 5). What stays shipped
and clean: consume-once root suspensions, the internally-parked quarantine +
laundering check, the flight recorder + beacons, the WSOD floor, the
pendingSleeps leak fix (confirmed by pendingSleeps=[] in the Nano dump). The
rare nested-rewind crash is ACCEPTED and fully observable until the
structural fix — the design-B fiber-first runtime (async/06,12,13), where
one scheduler owns every suspension and this interleave cannot exist.
.ci-cache-epoch 6→7.
Local: fiber 2/2 (one refusal beacon) + timer + firefox sweep 21 passed,
chromium scenarios 11 passed/4 quarantine-skips, web fatal+follow 2/2.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
move-vs-move diverged on CI right after value-vs-value was extracted: across
this week's runs, THREE different same-item conflict steps have each failed
to converge (settleConverged 90s, byte equality never reached). This is a
conflict-resolution class bug, not a per-step flake — same evidence profile
as before (no guard beacons, reproduces locally, not poll timing). Whole S4
test fixme'd alongside S4b; the sequenced-edit scenarios (S1–S3, S5–S8) have
never diverged and remain active. Tracking: memory s4-value-race-divergence.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
Two clients racing setValue genuinely diverge in ~5-8% of runs:
settleConverged times out at 90s with the trio never reaching byte equality.
Not timing (windows already widened, reproduces locally at single-worker) and
not the asyncify guards (zero beacons in failing runs) — a real CRDT/apply
race this harness exists to catch, gating unrelated releases in the meantime.
Extracted into its own test.fixme (S4b) with the full original body;
move-vs-delete and move-vs-move stay active in S4. Tracking notes: memory
s4-value-race-divergence — next steps are capturing both tabs' modelText diff
at timeout and bisecting the value-apply path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ac/code/session_019SE4o46Lnq3hF574FFq8x4
drift-trio-scenarios S4/S5 flaked 4× across this week's CI runs (three wasm
instances + full-suite load on the runner) while passing 77/78 locally under
stress — and the one local miss carried zero guard beacons, i.e. the same
under-load convergence shortfall, not a code path. Same treatment as the
follow spec: condition-based polls keep their shape, windows grow to what a
starved box actually needs (inline S4/S5 polls 20s→60s, waitAllContain
30s→90s, settleConverged 30s→90s). Local convergence stays ~1s.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
The first prod blue screen (finally!) also surfaced the console UX gaps: the
in-app log had to be hand-copied (and arrived truncated — the flight-recorder
dump missing), and the DOM-floor screen's log was a fixed block.
- React console: a copy button beside the toggle (writes the full log to the
clipboard, confirms in the log); stays collapsible on a fatal as in normal
editing.
- DOM-floor screen: its console mirrors the editor one — toggle bar
(▾/▸ console) + copy, collapsible — and its content is ring + trace dump.
- One blue family: both fatal screens now use the boot overlay's #1a1a2e, so
every full-screen state (boot, consent, fatal, floor) is coherent.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
The v0.1.23 flight recorder caught the kill live (console-export-2026-8-1_19-16-8):
dozens of benign fiber round-trips at w=0, the yield cycling healthily on its
buffer — then "fcs … ROOT w=1" and the trap, state frozen at Rewinding with
currData=root+20. The fatal condition, observed rather than inferred: a fiber
round-trip inside the ROOT's OWN sleep-wake continuation re-suspends and
re-rewinds the root nested inside its live wake rewind. Consume-once passed
correctly — it guards a different corruption and stays.
The round-3 deferral was aimed right but unscoped (taxed fiber-owned wakes,
flaked S4). Final form: every fresh sleep is tagged root- or fiber-owned
(fiber ⇔ started inside a finishContextSwitch fiber slice or a fiber-owned
wake; root entries don't count as slices); finishContextSwitch(root) defers
one macrotask ONLY while a root-owned wake is live (Asyncify.__wakingRoot).
Beacon: root-entry-deferred. Verified inert where it must be: zero beacons
across all 13 drift-trio-scenarios logs (26/26 + 25/26-then-26/26 stress —
the single miss carried no beacons, i.e. the pre-existing under-load flake).
Also: resume re-entries no longer push sleep contexts (the v0.1.23 dump
carried ~380 leaked zero-linked entries), and wake events in the recorder are
tagged R/f for ownership.
.ci-cache-epoch 5→6.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
The deferral (9ca2ac1) modeled the wrong condition and taxed every parked
fiber completion with a macrotask hop: under CI load that stretched
three-client apply chains and flaked drift-trio S4 twice consecutively
(26/26 green locally under stress) — retracted.
The actual fatal state, readable in all four prod stacks once seen: a SECOND
rewind of the same root suspension. Root suspends once per fiber_swap out of
it; two parked fibers completing against one root suspension epoch (a tool
fiber + a collab fiber both waking around open:settled) each drive
finishContextSwitch(root) — the second rewinds already-consumed data →
"unreachable executed" → poisoned runtime, with the wake-side "index out of
bounds" as the sibling symptom.
Cure: stop exempting root from the validity check the shim already keeps.
First consumption proceeds synchronously — zero added latency anywhere; the
second is refused ([wx-asyncify] "root suspension already consumed") — the
yielded fiber stays properly suspended and resumable, root continues via its
real pending resume, libcontext's ghost-epoch contract enforced one layer
lower. Root remains exempt only from the internally-parked quarantine (its
yield park is routine).
.ci-cache-epoch 4→5 (the epoch-4 cache holds the retracted deferral shim).
Local: fiber 2/2 + timer 1/1, firefox sweep 20 passed, drift-trio-scenarios
kicad-chromium 26/26 under 3-worker stress, web fatal+follow 2/2.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
v0.1.22's WasmErrorBoundary was still not enough: a commit-phase throw in
WasmTool's OWN effects unmounts the root, and no boundary below it helps.
fatal-screen.ts is the floor: plain-DOM blue screen with its own mirrored
log ring (append feeds recordFatalLog), installed at module import in
main.tsx — before and independent of React. It cooperates with the React
overlay: hidden while [data-testid="fatal-overlay"] exists, takes over via a
1Hz ensure-loop the moment it disappears. Fatal promotions also append the
asyncify flight-recorder dump so whichever screen survives carries the
targeting data.
fatal-overlay.spec.ts now also rips out the React root after the fatal and
asserts the DOM floor takes over with the mirrored [fatal] log.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
v0.1.22 still trapped with BOTH guards silent: the fatal rewind's target is
the ROOT context, which layer 2 exempted. All four prod stacks are the same
collision — a fiber completes its yield-back to main while main's sleep-wake
rewind is still on the stack (maybeStopUnwind → trampoline →
finishContextSwitch → doRewind(root) → unreachable), two "resume main" paths
interleaved in one tick; the 8ms-earlier "index out of bounds" is the wake
side of the same event.
Root entry is legal and constant in healthy flow; only the wake-window
overlap is fatal. So: serialize, don't refuse. The shim marks the
synchronous wake window (Asyncify.__inSleepWake around wakeUp) and DEFERS a
root finishContextSwitch landing inside it by one macrotask
([wx-asyncify] root-entry-deferred beacon, trampoline retry) — an ordering
change only, nothing dropped. Suspension recording happens before the
deferral branch, so the yielding fiber's validity survives the wake chain
nulling currData.
Plus a flight recorder: a 96-entry ring of asyncify/fiber events (sleeps,
wakes, every context switch with ROOT/wake-depth, refusals, deferrals),
silent in normal operation, auto-dumped with full machine state next to the
first trap signature in the console; window.__wxAsyncifyDump() on demand.
The next prod export reads like a black box, not a stack-shape puzzle.
.ci-cache-epoch 3→4 (wasm cache key omits scripts/**).
Local: fiber-resume-park 2/2 (one refusal beacon), timer-park 1/1, sweep 20
passed, web fatal+follow 2/2.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
4 flakes in 6 CI runs on 2026-08-01, all with clean traces: no guard beacons,
A demonstrably landed (the 208cb67 sequencing gate passed), B just never got
the one-shot rect applied within 30s while two wasm instances starved the CI
box. Awareness re-broadcasts state periodically, so a 90s condition-poll
converges on re-delivery; local runs stay fast (converge in <2s).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
kicadTestFiberParkStartSecond/PokeSecond: a second coroutine started while
the first body is asyncify-parked reproduces the misattributed jump that
launders the parked fiber past the C++ guard (the v0.1.21 prod bypass).
Spec scenario 2 stages it and asserts the JS stale-rewind guard quarantines
the laundered resume (exactly one fiber-resume-refused beacon), the parked
body completes undisturbed, and both coroutines finish cleanly.
Doc: async/16 rounds 2 + WSOD section.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
v0.1.21 still trapped with ZERO jump-refused beacons: the fatal swap PASSED
the C++ swap_suspended guard. Mechanism (async/16 round 2): a fresh JS entry
executing while g_current_context still points at a parked fiber gets
attributed to that fiber — fiber_swap writes a fresh, valid-LOOKING foreign
suspension into the parked fiber's struct and re-marks the flag. The flag
lies; the resume rewinds garbage.
This guard tracks truth at the emscripten-fiber layer (handlesleep.js wraps
Fibers.finishContextSwitch):
- valid suspensions = real swap-outs (currData == oldFiber+20 when the
trampoline runs), consumed on rewind;
- internally-parked = an entered slice that ended in a handleSleep park
(currData set, no nextFiber) — quarantined until a GENUINE swap-out,
where genuine means the fiber's pending sleep has resolved
(__pendingSleepContexts), so a laundering write cannot lift it;
- entering a quarantined or suspension-less fiber is REFUSED
([wx-asyncify] fiber-resume-refused, ghost contract).
The ROOT context is exempt from quarantine and refusal: its rewound
continuation runs the whole main loop, whose routine yield park says nothing
about a fiber body — the first build of this guard quarantined main off that
signal and starved every coroutine return (19 collab e2e reds, empty
results). Root = the old side of the first switch ever.
.ci-cache-epoch 2→3: the wasm output cache key omits scripts/**.
Red/green: fiber-resume-park.spec.ts scenario 2 (laundered resume → exactly
one refusal beacon, both coroutines complete); full fiber-heavy sweep green
(21 passed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
All three prod crash cascades (v0.1.19–21) ended white for the same reason:
the final trap lands inside a child's EFFECT (an embind call reached through
a react-query subscription), React unmounts the entire root, and the fatal
overlay + console panel die with the tree they were built to survive.
- WasmErrorBoundary: crash-capable children live inside it; the fatal screen
and the console panel live OUTSIDE and keep rendering. WasmTool state
(logs included) survives a descendant render/effect throw.
- The fatal screen is now an actual blue screen (:( + solid blue), and every
promotion path (window error, unhandled rejection, worker error, boundary)
auto-opens the console — the log is the only account of what was loading.
- tests/web/fatal-overlay.spec.ts pins the contract: a terminal uncaught
error after boot ⇒ visible fatal overlay + open console with the [fatal]
record. Green locally (23.9s).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
"B never tracked the second viewport move" ate three v0.1.21 release
attempts (2026-08-01) while the same build passed 10/10 locally, including
under 4-worker contention. The T2 block asserted B's convergence without
first confirming A's second fit LANDED — the T1 block has that sequencing
point, T2 didn't — so on a starved CI box any A-side fit delay was
misreported as a B-side tracking failure, and B's 30s window burned while A
was still moving. Mirror the T1 pattern: poll A onto its target, then start
B's clock.
Verified 4/4 green locally (web-chromium, 2 workers). Reminder that cost 20
minutes tonight: web/standalone/.env poisons local web e2e (untitled/title
mismatches) — stash it before local runs (standalone-env-e2e-poison).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
Companion to kicad f0ce20ef64 (libcontext swap_suspended guard), which this
pins. The v0.1.20 diagnostics decoded the crash that survived v0.1.13–19:
TOOL_MANAGER Resume()s a coroutine whose body is asyncify-parked inside
handleSleep, the swap rewinds the stale fiber suspension, and the runtime is
poisoned. Full chain of evidence in docs/features/async/16-fiber-resume-guard.md
(+ round-3 addendum in 15-timer-park-repro.md).
- wasm/bindings/fiber_park.h + kicadTestFiberPark{Start,Prime,Poke,State}
exports (pcbnew + merged kicad_editor): stages Call→yield→legitimate
resume→sleep park→mid-park Resume, the exact prod state machine. The
first yield matters: it primes a real (then stale) suspension, matching
long-lived tool loops rather than a first-slice park.
- tests/kicad/fiber-resume-park.spec.ts: asserts the healthy contract on
polled state only (embind returns across fiber swaps are unwind
placeholders). RED on the unguarded build — fiber/sleep buffer
cross-restores, a jump-ghost beacon, the parked body zombified. GREEN with
the guard: mid-park poke refused ([collab-fcontext] jump-refused beacon),
park completes, post-yield resume works, no trap signatures.
- Regression sweep green: timer-park-repro, collab-load-fuzz, load-pcb,
pcbnew-collab, collab-undo, eeschema-collab (19 passed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019SE4o46Lnq3hF574FFq8x4
