Squash of experiment/ff-big-modules vs main.
Big-module routing removed: native-EH shrank kicad_editor below
SpiderMonkey's x86-64 code budget (runs 29355049705/29356152413 green on
stock Firefox), so BIG_MODULE_SPECS routing and the baseline-only-JIT
crutch are gone — kicad-firefox and kicad-chromium both run the full
suite, with the module compiled the way real users' browsers compile it.
Per-engine screenshots end to end: stableShot/shotPath write
test-results/<engine>/<name>.png; baselines move to
baseline-screenshots/{chromium,firefox}/ and the whole tools/screenshots
pipeline (compare/promote/manifest/spec-map/changelog/Discord) keys on
<engine>/<name>. Previously Firefox and Chromium renders of one spec
overwrote each other and Firefox renders were never actually gated.
Seeded from CI run 29421380806 (92 new firefox baselines, +24 chromium
web-suite shots); manifest generated from the baseline tree.
One merged playwright.config.ts (kicad/asyncify/coroutine/perf as
projects); ~25 dead npm scripts dropped. The web suite is gated in CI for
the first time ever (4 rotted specs fixed, 5 broken lib-bridge specs
triaged as fixme in docs/features/web-e2e-rot/); cheap lint step after
npm ci; last 26 blind-sleep violations fixed.
SwiftShader retired: CI Chromium renders WebGL on ANGLE → Mesa llvmpipe
(--use-gl=angle --use-angle=gl --ignore-gpu-blocklist; the blocklist flag
is mandatory — llvmpipe is blocklisted and WebGL is silently unavailable
without it) in BOTH configs. Under WORKERS=4 congestion SwiftShader
transiently failed the first post-board-load draw and the recovery
cascade ended in a silent permanent Cairo fallback — that engine flip was
the "~1.2% changedRatio both directions" occ-export baseline flake.
Validated 160/160 across two 80-repeat rigs; full analysis in
docs/features/wx-parity-bugs/occ-export-context-eviction.md. Chromium
baselines shift slightly on llvmpipe — promote once from the first green
run. Deflakes the new coverage exposed: presence baselines settle before
capture; presence fixtures declare current file formats; perf gets its
own outputDir so CI evidence survives; occ-export settles the board paint
before the export dialog; menu-item waits (waitForRenderedByLabel before
clickMenuItem) in 4 specs + the TESTING.md rule.
Web suite runs the PROD build, in parallel: webServer becomes backend
`start` + the standalone's e2e:preview (build-preview.mjs: link-wasm →
stash the public/wasm symlink aside during vite build, build-demo.mjs's
move — then vite preview as the persistent server). The wasm middleware
serves /wasm/* in preview and emits COOP/COEP/CORP itself (a pthread
worker script's own response must carry COEP or Chrome kills it with
ERR_BLOCKED_BY_RESPONSE). VITE_* flags bake at build time;
VITE_ALLOW_USER_OVERRIDE joins turbo globalEnv. fullyParallel + default
workers: 5.2m → 1.4m. Determinism fixes the parallel run exposed:
shared-page specs become serial groups; locks.spec grabs alice's exact
item via the new kicadCollabTestSelectByUuid hook (cross-tab "first
footprint" order is not a ysync invariant); quit specs poll page.url()
(quit supersedes its own navigation — NS_BINDING_ABORTED on Firefox).
Suite: 51 passed / 12 skipped / 0 failed in 1.6m.
CI-coverage gate (lint:ci-coverage): every tests/**/*.spec.ts must be
reachable from the npm scripts the workflows invoke — scraped from
.github/workflows/, resolved through package.json, coverage asked from
playwright --list itself. Rules: uncovered-spec + orphan-project (with a
documented LOCAL_ONLY_PROJECTS allowlist). Gating next to
lint:determinism; 138 spec files / 13 projects accounted for.
Product fixes kept from the investigations (reachable on real GPUs too):
wx 7799fd1be5 — paint flags clear before dispatch + Invalidate always
propagates; kicad 3dcfea5e45 — SwiftShader pass-boundary flush +
per-instance font texture + first-frame GL-error drain (GAL recovery
recovers instead of falling back to Cairo) + the user-facing eeschema
switch navigates again under __EMSCRIPTEN__ (project-sync's
FaceRegistered gate had rerouted it into the hidden sync player; caught
by the newly-gated web suite).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018eUxiPApHgGiu9NFyQfhAq
Make the Playwright e2e + kicad suites deterministic so screenshot flake stops
tracing to timing races.
- Blind page.waitForTimeout -> condition waits (expect.poll, web-first
assertions, waitUntil) + readiness helpers (waitForWxApp, waitForCanvasApp).
Remaining sleeps are documented interaction dwells (annotated).
- Defensive "if element exists" branches -> loud asserts; label-fallback chains
-> normalized clickMenuItemByText. First-run wizard for/if loops removed by
seeding calculator/gerbview/pcbnew HTMLs.
- Screenshots: new stableShot(page, name) settles the render in-page (canvas
hash over rAF) then writes a raw PNG to test-results/ for the existing offline
gate (tools/screenshots vs baseline-screenshots). Replaces toHaveScreenshot,
which did inline compare + its own baselines and had decoupled the specs from
the real gate. scale:'css' pinned.
- retries: 0 in both configs.
- Guard: tests/tools/lint-determinism.ts (npm run lint:determinism) bans blind
sleeps / toHaveScreenshot / inline retries / swallowed catches in specs;
documented exceptions carry a marker. Rules in tests/TESTING.md.
Assertions, coverage, and renders unchanged (semantic-equivalence reviewed;
captures pixel-identical modulo inherent timer/timestamp/3d-raytrace variance).
Both suites green at retries:0 (e2e 340, kicad 92); ~35-61% faster.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BVX1pHMvRPYHdp6ZfEawrk
The 3d-webgl merge (kicad eb13ff3bdc: the viewer now defaults to the real
OpenGL renderer via wasm/gl1, and occ-split moves STEP parsing into the
occ_service worker) made the raytracer-era orchestration on this branch moot —
main's chromium-ci phase is green at 15-way parallelism (28666407570 /
28698861536). Drop what no longer earns its complexity, keep the diagnostics,
fix main's live flake, and make the deadlock spec test what it was written for.
- REVERT the chromium-ci-3d serial project, the two-phase test:kicad:ci, the
SwiftShader GPU-process flags, and the resize-drag/models skips: config and
package.json are byte-for-byte back to main's shape. The raytracer contention
they guarded is no longer on the CI path.
- FIX main's live flake: run 28698861536 is green only via retry
(3d-viewer.spec:26 flaky) and 28666407570's deadlock red sampled an ALL-ZERO
pixel signature — the viewer's first frame lags the canvas's creation on
software WebGL under parallel load, and sampling too early reads an all-black
backbuffer. New waitForThreeDRender() gates render assertions on actual
pixels (1s-interval full-frame CPU reads) instead of fixed sleeps, used by
3d-viewer.spec:26 and the models render tail.
- KEEP the storm-proofed samplers (one full-frame getImageData on a
willReadFrequently canvas replacing 256 per-pixel GPU round-trips per sample
— the "GPU stall due to ReadPixels" trigger) and the logThreeDDiag
instrumentation: engine-independent, and they de-risk every remaining
software-GL pixel read.
- models spec: bridge assertions stay front-loaded (the protocol regression
signal is independent of the render); the occ_service parse verdict is now
POLLED — it lands async relative to the bridge ensures, so asserting it
immediately raced the worker; the render tail runs again everywhere. (The
pre-webgl raytracer+models renderer-death documented in a17f3be does not
affect the OpenGL default path — the raytracer-toggle+models combination
remains untested product surface, tracked outside this branch.)
- deadlock spec: the deadlock it guards is raytracer-specific and the viewer
now defaults to OpenGL — on the GL engine it either passes vacuously (fast
renders make every liveness assertion trivial, 28698861536) or fails on the
black first frame (28666407570). It now flips the engine via the "Use
raytracing" toolbar toggle (loud assert if the toggle moved) and
cross-checks engagement by requiring the canvas pixels to CHANGE after the
flip with no input in between (the raytraced frame is lit differently; a GL
re-render reproduces identical pixels; heap growth is unusable — mimalloc
satisfies the raytracer from freed arena pages). That guard immediately
caught a REAL defect: on the webgl-era wasm build the toggle is INERT (the
click lands and "Reload time" updates, but the canvas never changes —
suspects: DoRePaint's silent catch(runtime_error) freezing the canvas after
a raytracer Redraw throw, or ToggleRaytracing writing m_boardAdapter.m_Cfg
while RenderEngineChanged() reads GetAppSettings<…>(), possibly different
instances in the merged bundle). The spec is therefore test.skip-annotated
as a KNOWN ISSUE with the full engine-force machinery in place — unskipping
it self-validates the product fix. The CI-skip also stays (raytracer
liveness needs real-GPU pacing; the Worker-boot deadlock mechanism is
covered on CI by the standalone wx harnesses).
- 180s viewer-open waits kept as pure CI headroom (never slow a passing run).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The headed Mesa-llvmpipe switch (ce44636) was built on a misdiagnosis. With
--use-gl=angle --use-angle=gl under Xvfb, Chromium gets NO WebGL AT ALL on the
GPU-less CI VM: runs 28652367347 + 28664038296 both log
`glcanvas count before opening 3D viewer: 0` (even the pcbnew board's own GAL
canvas fails to create) and every failure is the FIRST viewer-open wait timing
out. "Timeouts instead of crashes" meant no GL context existed to crash, not
that llvmpipe was stable-but-slow. The Mesa/GLX path works for the Firefox
project, but not for Chromium on this image.
Headless SwiftShader is the environment where GL provably works — the
serialized run 28649537489 opened the viewer and rendered/sampled 56 colours
(viewer:26/130/209 all passed single-tab). Its remaining failures split into
two families, each addressed on its own evidence:
1) SwiftShader GPU-process kills (deadlock spec's lost canvas/black frame):
- chromium-ci-3d back to headless + --enable-unsafe-swiftshader, plus
--disable-gpu-watchdog and --disable-gpu-process-crash-limit: heavy churn
stalls the software-GL GPU process until the watchdog shoots it
(CONTEXT_LOST -> removed GL canvas -> "Target crashed"); with the watchdog
off a slow op completes, and without the crash limit a lost GPU process
keeps restarting for later tests. Same args locally (no-ops on a healthy
real GPU).
- Storm-proof every canvas sampler (3d-viewer, models, deadlock specs): one
full-frame getImageData on a willReadFrequently (CPU-backed) 2D canvas +
JS grid sampling, replacing 256 per-pixel getImageData GPU round-trips per
sample — the "GPU stall due to ReadPixels" trigger, worst in the deadlock
spec's 1.5s-interval settle polling. Sampled pixels and sig order are
unchanged.
- edge-resize spec: keep the open + 5-resize-handles assertions on CI, then
test.skip the drag — each resize step re-raytraces synchronously on the
wasm main thread and the 12-step drag blocked mouse.move past the 240s
budget even single-tab. Real-GPU-paced; still runs locally (deadlock
spec's CI-skip from d9900fe unchanged, same rationale).
2) KNOWN PRODUCT BUG uncovered in the models spec (NOT SwiftShader, NOT CI):
raytracing a scene WITH component models kills the Chromium renderer
process outright ~6s after scene build — silent process death with a FLAT
wasm heap (531-637 MB watched at 2s intervals; not OOM), no console error,
no wasm abort, no crash report. Deterministic on a real GPU (5/5),
fixture-independent (700 KB USB-C and 61 KB USON-8 STEP crash identically);
the same death hits CI SwiftShader at ~66s. Board-only raytraces complete
and stay alive, so the bug is specific to the model path. The spec's one
historical green ended at raytrace-age ~6s — inside the death window by
luck; every CI run of it ever died. Fix here: front-load the bridge
assertions (ensure requests fire during scene BUILD, before the raytrace),
so protocol regressions still fail the suite everywhere, and test.skip the
render tail unconditionally with the bug documented at the skip site. The
raytrace->blit pipeline stays gated on CI by 3d-viewer.spec's board-only
render. Bug tracked for a dedicated fix session.
Correct the now-wrong llvmpipe comments (config, open-wait rationale, deadlock
skip reason). Expected CI matrix: viewer open/render/stacking/titlebar +
models-bridge on SwiftShader serial; edge-resize drag + camera-move deadlock
on real GPUs locally; the Worker-boot deadlock mechanism stays covered on CI
by the standalone wx harnesses.
Locally validated on a real GPU (--project=chromium): 5 passed + 1 skipped
(the documented models render tail).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Follow-up to the headed-llvmpipe switch. That run (28652367347) stopped CRASHING —
every failure became a plain timeout, confirming llvmpipe gives a stable WebGL context
(no more CONTEXT_LOST / "Target crashed") but software raytracing is just slow:
- The first board raytrace takes ~60s on llvmpipe, right at openThreeDViewer's 60s
wait → raise it to 180s (shared helper + the models spec's inline copy). Real GPU
returns in ~2s, so it is only CI headroom.
- The deadlock spec asserts the wasm main thread stays responsive within ~15s DURING a
raytrace — a premise only true on a fast (real) GPU. Under slow software rendering a
legitimately-slow raytrace is indistinguishable from a deadlock, so skip it on CI
(test.skip on process.env.CI). The Worker-boot deadlock MECHANISM stays covered on CI
by the standalone coroutine-pthread-ondemand / raytrace-threads wx harnesses; the spec
still runs locally on a real GPU.
- Gate the headed + llvmpipe launch flags on CI so chromium-ci-3d behaves like a normal
bundled-Chromium project locally.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The three heavy 3D-viewer specs (3d-viewer, 3d-viewer-deadlock, 3d-viewer-models)
crashed on CI (run 28604015154): "Target crashed", "browser has been closed", and a
black frozen canvas. Each boots the 3D-enabled pcbnew build, which pre-warms
~hardwareConcurrency*2+8 (~68 on the 30-core VM) Web Workers AND runs a multi-threaded
CPU raytracer over SwiftShader software-WebGL. Under fullyParallel the three spec FILES
run in separate browser processes concurrently, and the simultaneous raytrace tabs
exhaust the shared GPU process's ~16 live-WebGL-context limit / renderer memory. The
pool pre-warm (the deadlock fix, 7630c7e) is correct and present in the tested binary;
per-file isolation prevented in-process Worker accumulation but NOT this cross-process
concurrency — the gap these specs' first CI exposure hit.
Fix (test-orchestration only; no wasm/submodule changes):
- Route the three specs to a dedicated `chromium-ci-3d` Playwright project (own
outputDir=pw-artifacts/kicad-3d so its failure traces survive the later perf run's
wipe; the existing tests/pw-artifacts/** upload already captures it).
- test:kicad:ci runs it as a SECOND `playwright test` pass with --workers=1, so at most
one raytrace tab is alive at a time (the environment in which a single 3D load is
reliably green — docs/features/3d-raytracer/README.md §7). An exit-code-aggregating
sequencer runs both phases and fails if either fails (no silent skip of the 3D pass).
- Cap phase-1 CI workers to 12 as headroom under the WebGL-context cap.
- Log a resource-diagnostic snapshot (hardwareConcurrency, pthread pool, GL-canvas
count, wasm/JS heap) before each 3D interaction so any recurrence is attributable.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Viktor Vaczi