Windows users with the OS temp dir on a small system drive have hit
C: exhaustion mid-render (Slack ts=1784219488 · CLI v0.7.58 · win32
15 GB / 8-core, ~5500 frames). The engine already honors
HYPERFRAMES_EXTRACT_CACHE_DIR for relocation, but the knob was
undocumented and invisible in diagnostics — the reporter had to piece
together a 4-flag compound workaround including EXTRACT_CACHE_DIR=off.
Changes:
- Extract the env-var resolver into a public engine API
(resolveExtractCacheDir, defaultExtractCacheDir,
EXTRACT_CACHE_DIR_DISABLED_ALIASES) with a typed resolution shape
distinguishing "disabled by user" vs "default" vs "env override".
- Add a Frames-cache check to `hyperframes doctor` that reports the
effective directory, its free space, source (env or default), and
fails with a relocation hint when <2 GB free at that mount.
- Add `hyperframes render --frames-cache-dir <path>` as discoverable
CLI sugar for the env var, including the opt-out aliases
(off/none/false/0) and CWD-safe absolute-path resolution.
- Document the flag in docs/packages/cli.mdx with the field-signal
citation, and add a render example row for the Windows workflow.
- Cover both surfaces with unit tests (6 doctor cases + 4 engine
cases including all disabled-alias variants).
Refs Slack #hyperframes-cli-feedback ts=1784219488 (win32 v0.7.58).
Co-authored-by: Via <via-heygen[bot]@users.noreply.github.com>
Field-signal baseline: >=2 fallbacks/hr on darwin/arm64 from filter:blur
and filter:drop-shadow triggers. Fallback path perf is currently untimed,
so we can't know if the overhead is 10% or 10x. This PR adds opt-in
per-frame timing (HF_PROFILE_FALLBACK_CAPTURE=true) that emits p50/p95/p99
+ trigger reason via the observeRenderStage telemetry channel extended in
#2510. Diagnostic surface only -- no perf fix, no behavior change on
healthy paths.
Stack: PR #9 (final) of 9 (base via/escape-hatch-fallback-reproducer).
Signed-off-by: Via
Field signals ts=1784049136 (hardware-GPU intermittent black rectangles →
resolved with --no-browser-gpu --low-memory-mode --workers 1) and
ts=1784032286 (clip-path animated image → intermittent black rectangles →
resolved with deterministic precompose). Pattern: hardware-GPU writes
solid-black on some composition shapes; software-GPU / screenshot bypass
restores correctness. Raw per-pixel diff alone false-positives on every
compositor jitter frame; the diagnostic-grade signal is asymmetric
black-only-in-A pixels (solid-black where B has content).
Adds `packages/engine/src/utils/gpuParityDiff.ts`: pure helpers
(`diffGpuParityFrames`, `diffGpuParityPngs`, `verifyGpuParity`) that
compare two RGBA frames captured via different GPU paths, count per-pixel
diffs above a tolerance, and isolate black-only-in-A / black-only-in-B
pixel counts + bounding boxes. Symmetric black regions (real black content
present in both captures) are NOT flagged. PNG wrapper preserves the
underlying decode error as Error.cause on either side. All exposed via
`@hyperframes/engine`'s package index for downstream wiring.
19 unit tests cover identity, per-pixel tolerance, the field-bug shape,
the shared-black no-op case, bounding-box tightness across multiple
regions, the inverse pattern, dimension mismatch, data-length mismatch,
overlapping threshold rejection, custom tolerance, verdict output, PNG
end-to-end, and cause-preservation on both A and B decode failures.
Reduced-scope first pass. Wiring a `hyperframes verify-gpu-parity` CLI
command, dual-mode capture orchestration, and integration coverage against
a known-bad composition is intentionally deferred to a follow-up so the
diagnostic primitive can land and be exercised in isolation. The exported
surface is stable — a follow-up need only add the capture-and-diff driver.
Stack: PR #7 of 9 (base via/parallel-capture-observability).
Signed-off-by: Via
Field signals ts=1784019503 (heartbeat reports 0 frames during 64s
browser calibration — reads as broken but is healthy) and ts=1784042064
(1292s Windows render hard-exited during video frame extraction with
no final error string — silent worker crash).
Add calibrating/capturing state to heartbeat labels; surface synthetic
terminal error on unexpected worker exit when no explicit error was
emitted.
Stack: PR #6 of 9 (base via/overlay-count-lint).
Signed-off-by: Via <vance@heygen.com>
Field signal ts=1784146416 (darwin/arm64, CLI 0.7.58, 7/10): host
page.goto hit Navigation timeout of 60000ms twice on a CSS 3D + audio
composition; Docker rendered the same composition successfully.
Puppeteer's stock "Navigation timeout of 60000 ms exceeded" text names
none of HyperFrames' existing escape hatches, so the reporter had no
signal that the failure had knobs.
Wraps main-render Puppeteer `page.goto` errors matching
/Navigation timeout|net::ERR_TIMED_OUT/i with an augmented message that
names:
- The effective timeout currently applied (`cfg.pageNavigationTimeout`).
- Raise-the-timeout: `PRODUCER_PAGE_NAVIGATION_TIMEOUT_MS` env,
`--browser-timeout` CLI flag (seconds).
- Browser-binary escape hatch: `HYPERFRAMES_BROWSER_PATH` env.
- Field-signal shape: darwin/arm64 + CSS 3D + audio compound Docker
hint — gated on all three inputs being explicitly true; falls back
to generic hints when any input is unknown.
Mirrors #2443's HYPERFRAMES_BROWSER_PATH surfacing pattern (which
covered download-time failures) at the runtime `page.goto` layer.
Non-matching errors flow through unchanged. Original error preserved
via `err.cause`.
Wired into `renderOrchestrator.executeRenderJob`'s top-level catch,
composed after `augmentProtocolTimeoutError` so the two augmenters
never both fire on the same error (mutually exclusive regexes).
Current wire-up passes no `hasCss3D` / `hasAudio` context — no
compile-time CSS-3D signal is threaded through the render pipeline,
and `hasAudio` is block-scoped inside the try. Per the helper's
fallback docs, unknown flags route to the generic env + browser-path
hints. A future compile-time CSS-3D scan can thread both flags to
enable the full compound Docker hint without touching this helper's
signature.
Stack: PR #3 of 9 (base via/win32-streaming-encode-autodisable).
Signed-off-by: Via <vance@heygen.com>
Field signal ts=1784131903 (win32/x64, CLI 0.7.58, 156s UI-heavy):
stable ONLY with four flags together — --workers 1 --no-browser-gpu
--low-memory-mode + PRODUCER_ENABLE_STREAMING_ENCODE=false. Since
--no-browser-gpu and --low-memory-mode already imply screenshot
capture, three of the four flags are structurally coupled. Auto-detect
the compound at resolveConfig time and disable streaming-encode on
the caller's behalf; user explicit-set (PRODUCER_ENABLE_STREAMING_ENCODE
or overrides.enableStreamingEncode) always wins.
Composition duration is not known at the config layer, so the wire-up
passes compositionDurationSec:undefined and the helper reduces to the
three-condition compound (platform + softwareGpuForced + workers=1).
The 4-arg helper stays exported for downstream callers that DO know
duration (e.g., renderOrchestrator) and want the >120s guard.
Trade-off documented in code + PR body: false positives possible for
short (~<120s) Windows software-GPU single-worker renders. Mitigation
is the explicit opt-in escape hatch.
Emits a single [hyperframes] log line naming the trigger + how to opt
back in, so operators can tell an auto-disable apart from an explicit
opt-out. Adds streamingEncodeAutoDisabledOnWin32Compound internal
provenance for downstream telemetry.
Stack: PR #2 of 9 (base via/protocol-timeout-discoverability).
Signed-off-by: Via
Field signal ts=1784047847 (darwin/arm64, 8GB M1, 9 videos + 22 images):
reporter hit Runtime.callFunctionOn timeout and switched to FFmpeg
because the error didn't surface HyperFrames' existing knobs
(PRODUCER_PUPPETEER_PROTOCOL_TIMEOUT_MS env, --protocol-timeout CLI).
Wraps main-render Puppeteer errors matching /Runtime\.callFunctionOn
timed out|Target closed|protocolTimeout/i with an augmented message that
names the effective timeout, the env var, the CLI flag, and the
field-signal shape. Non-matching errors pass through unchanged
(returned as the same instance). Original error preserved via err.cause.
Also adds a dedicated --protocol-timeout row to the CLI docs Flags table
so PRODUCER_PUPPETEER_PROTOCOL_TIMEOUT_MS is discoverable via search.
Signed-off-by: Via <noreply@heygen.com>
Deepwork's request-changes on #2411 (twice): deFallbackReason's blank/psnr
split still ran /blank/i.test(err.message) even after this PR's stated goal
of moving off message-text parsing — a reworded message, a translated
string, or a differently-shaped error crossing a module boundary could
silently relabel a blank failure as psnr (or vice versa), corrupting the
soak's telemetry taxonomy.
DrawElementVerificationDetails now carries a required `kind: "blank" | "psnr"`
field, set at all three real throw sites in captureStreamingStage.ts. The
orchestrator derives deFallbackReason from getDrawElementVerificationDetails's
kind instead of regexing the message. Making `kind` a required constructor
argument means any future throw site that omits it fails to compile, closing
the gap for good rather than just at today's three call sites.
New tests in frameCapture.test.ts prove message-independence directly: kind
survives a reworded message that says neither "blank" nor "psnr", and stays
correctly "psnr" even when the message adversarially contains the substring
"blank" — the exact scenario a regex-based classifier would get wrong.
de_fallback_reason only told you the fallback happened (blank/psnr/oom/
capture_error), not the failing PSNR or frame index — that data existed as
text inside the thrown error's message and was discarded on the way to
telemetry. DrawElementVerificationError now carries structured
frameIndex/failedDb/verifyThresholdDb; the orchestrator reads them via the
new getDrawElementVerificationDetails helper instead of regexing message
text, and both telemetry surfaces (the render_complete perfSummary path and
the crash-survival RenderCaptureObservability mirror) emit
de_fallback_failed_db / de_fallback_frame_index.
Needed to distinguish "32dB vs the 32dB threshold, tune it" from "12dB real
corruption, investigate" during the parallel-router soak — currently that
distinction is invisible.
## What
- cap static-dedup verification at 15 seconds of wall-clock time
- disable the optimization and continue normal capture when the verification budget is exhausted
- add a regression that models the reported 350-second / ~8,400-frame alpha render
## Why
Static-frame verification uses full-page screenshots. Its existing screenshot-count budget still scales with composition duration, so a long composition can spend minutes proving an optimization before frame capture starts. The reported 350.35-second ProRes alpha render spent about eight minutes in this phase before safely disabling dedup.
## How
The verifier now records a deadline before seeking verification frames. It checks the deadline before every full-page capture and returns the existing fail-closed `budgetExhausted` result when time is exhausted. This keeps the existing density-based safety checks while bounding their startup cost.
## Test plan
- [x] Unit tests added/updated
- [ ] Manual testing performed
- [ ] Documentation updated (not applicable)
- [x] Focused engine test: 9/9 passed
- [x] Engine typecheck passed
- [x] Pre-commit lint, format, tracked-artifact, fallow, and typecheck gates passed
- [x] Full engine suite: 988 passed, 3 skipped; 2 pre-existing environment failures because host FFmpeg 4.2 lacks `-fps_mode`
Miguel R4 blocker on #2359: my R3 fix at renderOrchestrator only updated
the observability copy, leaving the authoritative captureForceScreenshot
local at compileResult.forceScreenshot (false for auto→software). The
frameCapture side clamped its own local and correctly routed screenshot,
but downstream orchestrator code overwrote observability back to
beginframe from the still-false local at two sites:
- Parallel-stream label at renderOrchestrator.ts:2293 mis-labelled the
stream as 'beginframe' when actual capture was 'screenshot'.
- capture_strategy telemetry at renderOrchestrator.ts:2440-2450
overwrote the earlier observability correction, so the final
captureMode observation flipped back to 'beginframe' while the
engine actually captured screenshot.
Fix: extract the clamp into a caller-facing helper
applyConcreteGpuScreenshotClamp(current, resolvedGpuMode, cfg) that
returns the (possibly-promoted) new boolean. Callers assign it back to
their authoritative local, so routing + telemetry + strategy code read
one value.
Changes:
- packages/engine/src/config.ts: new exported
applyConcreteGpuScreenshotClamp; delegates to
shouldClampToScreenshotForConcreteGpu but computes the caller's
final value, not just the clamp decision. Reads the programmatic
opt-out from cfg.forceScreenshotExplicitlyOptedOut. Idempotent on
already-true input.
- packages/engine/src/index.ts: export the new helper.
- packages/engine/src/services/frameCapture.ts: replace the inline
OR expression with applyConcreteGpuScreenshotClamp.
- packages/producer/src/services/renderOrchestrator.ts: assign result
into the AUTHORITATIVE captureForceScreenshot local (was updating
only observability). Downstream parallel-stream label at :2293 and
capture_strategy telemetry at :2440-2450 now read the corrected
value.
Tests: 6 new caller-level cases for applyConcreteGpuScreenshotClamp
covering the exact matrix Miguel called out:
- resolved software + default false → promotes to true (screenshot)
- resolved software + programmatic opt-out → stays false (BeginFrame)
- resolved hardware + default false → stays false
- resolved software + already-true → stays true (idempotent)
- resolved software + env PRODUCER_FORCE_SCREENSHOT=false → stays false
- resolved software + undefined cfg → promotes to true (frameCapture path)
Local: 67/67 engine config tests pass (was 61). oxfmt clean.
Miguel R3 blocker on #2359: the runtime helper only checked the env opt-out
(PRODUCER_FORCE_SCREENSHOT=false), silently defeating the documented
programmatic escape hatch (overrides.forceScreenshot === false) on the
browserGpuMode:'auto' → software probe path. At the concrete-resolution
site the boolean forceScreenshot === false is ambiguous between default
and explicit opt-out — resolveConfig sees the provenance but the runtime
helper does not.
Fix: persist provenance on the resolved config.
- New INTERNAL EngineConfig field forceScreenshotExplicitlyOptedOut, set
by resolveConfig when EITHER env or programmatic explicit-false is
present. Purpose-documented in the type as 'not intended to be set by
callers'.
- shouldClampToScreenshotForConcreteGpu gains an opts.programmaticOptOut
parameter; returns false early when set. Env stays as the third arg
(backward compatibility with existing tests).
- frameCapture.ts and renderOrchestrator.ts pass
config.forceScreenshotExplicitlyOptedOut through at both call sites, so
the auto→software probe path preserves the same escape hatches as
literal browserGpuMode:'software'.
New tests: 5 additional cases across the helper (programmatic opt-out
alone; programmatic beats missing env) and resolveConfig provenance
(programmatic sets flag; env sets flag; neither leaves it undefined).
Local: 61/61 engine config tests pass (was 56).
The SDR-to-HDR extraction tests synthesize their HDR fixture with
-color_trc/-color_primaries flags and rely on the encoder propagating
them into the bitstream. The pinned Windows CI ffmpeg build drops the
transfer on that path, so after #2377 narrowed HDR detection to the
transfer function, the fixture probes as SDR on Windows and both tests
fail (hdrPreflightCount 0). Write the VUI directly with the
h264_metadata bitstream filter so the tag survives on every build.
Addresses Miguel's R1 blockers:
1. `browserGpuMode: "auto"` that runtime-probes to software slipped past the
`resolveConfig` clamp — that clamp only sees the pre-resolve string. Add
`shouldClampToScreenshotForConcreteGpu(resolvedGpuMode, currentForceScreenshot, env)`
in `packages/engine/src/config.ts` and apply it at BOTH concrete-resolution
sites:
- `packages/engine/src/services/frameCapture.ts`: downgrades `preMode`
from "beginframe" to "screenshot" when resolved GPU is software (respects
`PRODUCER_FORCE_SCREENSHOT=false` env opt-out), fixing the routing.
- `packages/producer/src/services/renderOrchestrator.ts`: updates
`captureObservability.forceScreenshot` (and thus `captureMode`) at the
same call site, fixing the observability truth on the auto → software
case.
2. New unit tests in `config.test.ts`:
- Documents the auto-branch gap (resolveConfig leaves auto as
forceScreenshot=false — the runtime companion closes it).
- 5 branch tests on `shouldClampToScreenshotForConcreteGpu` covering
software / hardware / already-forced / env-opt-out / non-"false" env
values.
Full suite: 56/56 pass.
Scope narrowing on Blocker 2: the distributed rendering path at
`packages/producer/src/services/distributed/plan.ts:753-754` and
`renderChunk.ts:462-466` explicitly hardcodes `browserGpuMode:"software",
forceScreenshot:false` post-resolveConfig and stays outside this PR's
invariant boundary. `compileStage` may still flip it to true for alpha
formats, but generic MP4 distributed renders on SwiftShader hosts remain
BeginFrame. That's a separate architectural cleanup (needs its own
behavior-change trace); the PR body now scopes the invariant to the
in-process CLI/orchestrator path.
When `browserGpuMode === "software"`, set `forceScreenshot = true` in
`resolveConfig`. Explicit opt-outs (`PRODUCER_FORCE_SCREENSHOT=false`
or `overrides.forceScreenshot === false`) are honored.
This is defense-in-depth on top of the existing platform gates:
1. Linux + software (SwiftShader host) skips BeginFrame, avoiding the
compositor stall on shader-heavy frames under CPU raster (same
motivation as the closed PR #822).
2. `renderOrchestrator`'s reported `captureMode` field is derived from
`cfg.forceScreenshot ? "screenshot" : "beginframe"` — without this
clamp it misreports `"beginframe"` for the actual screenshot capture
on darwin + software.
3. Any new BeginFrame or drawElement entry point that forgets to gate
on GPU mode still routes to screenshot here.
Does NOT fix SwiftShader-on-darwin text-rasterization artifacts (an
ANGLE-SwiftShader issue on macOS text — the fix there is to use
`--browser-gpu`, which routes to `--use-angle=metal`).
Three defects found by max-effort code review of this branch:
1. The Bun OOM exact-match regex was defeated by this codebase's own
parallel-worker error wrapping. executeParallelCapture/formatWorkerFailure
(parallelCoordinator.ts) always wrap a worker's error as
"Worker N: <message>", optionally suffixed and joined with other workers'
segments, all prefixed "[Parallel] Capture failed: ". That wrapping
defeated the exact-message check for exactly the cohort (deParallelRouter
routed, N separate Chrome processes) the OOM-drops-to-1 fix targets — a
real OOM there would retry at the SAME worker count instead of dropping
to 1. Added a second pattern that recovers the signal by requiring
"out of memory" appear as the WHOLE content of a "Worker N: ..." segment
(bounded by end-of-string/"; "), preserving the same exact-match property
(no bare substring match) while surviving the wrapping. Verified against
the real wrapping logic, not a hand-typed guess at its shape.
2. shouldRetryViaPinnedFallback didn't exclude cancellation, so aborting a
render mid-capture on the pinned router/inversion cohort would detour
through spawning a fresh encoder/capture session before the outer catch's
RenderCancelledError branch ended the render — delaying "stop" with a
pointless resource spin-up/tear-down. Added an isCancellation param
(checked first, before isVerifyError) using the same
`err instanceof RenderCancelledError || abortSignal?.aborted` check the
outer catch already uses.
3. deFallbackReason (this PR's new "oom"/"capture_error" values) was set
locally but never mirrored into RenderCaptureObservability alongside
deSelfVerifyFallback, so a render that fails AFTER a fallback attempt
(perfSummary never built) was indistinguishable in render_error telemetry
from one that never attempted any fallback — undercutting the "how often
does the OOM retry fire on a render that still ultimately fails"
question this branch exists to answer. Threaded through
RenderCaptureObservability → RenderObservabilityTelemetryPayload →
renderObservabilityTelemetryPayload, mirroring the existing
deSelfVerifyFallback plumbing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found while testing the previous commit's OOM-drops-to-1-worker fallback
end-to-end: the producer's deployed runtime is Bun (JavaScriptCore), not
Node (V8) — see packages/gcp-cloud-run/Dockerfile's `bun dist/server.js`
entrypoint. All 7 MEMORY_EXHAUSTION_ERROR_PATTERNS are V8-specific allocation
failure signatures; JSC's equivalent for the same single-oversized-allocation
RangeErrors is the bare string "Out of memory" (verified against real Bun
behavior), which none of them match. Without this, isMemoryExhaustionError
returns false for genuine production OOM, so the memory-specific worker-count
reduction just added would never actually engage where it's deployed — every
OOM would fall through to the generic capture_error retry path instead.
Matches the FULL (trimmed) message only, not merely a substring — same
rationale as the existing V8 patterns' comment: "out of memory" also appears
in benign WebGL/GPU console noise that must not trip this classifier.
Verified end-to-end from a script inside the producer workspace (importing
the real @hyperframes/engine source, not a stale globally-cached npm dist a
script outside the workspace would otherwise resolve to): a genuine Bun
RangeError from new Uint8Array(Number.MAX_SAFE_INTEGER) now correctly
classifies as memory exhaustion and drives both resolveInversionRetryPlan
and resolveParallelRouterRetryPlan down to workerCount=1 on retry.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Address PR #2093 review feedback (Miga, Rames D Jusso):
- The walker treated a repeating nested timeline (total > single) as an
opaque interval and never descended into it, so a tl.call() living
inside one would slip past hasTimelineCall detection entirely — the
"any tl.call() disqualifies" claim wasn't quite literal. Now recurses
for detection purposes even when the span is already opaque; the
parent-level interval still dominates for frame-animated-marking, so
this only widens what counts as "has a call()," never narrows the
existing interval coverage.
- Restored the totalDuration() vs duration() rationale comment that got
dropped when the tl.call() detection comment was added above it.
Real bug report: a mono count span driven by a GSAP tl.call() (a counter
going "0 sur 0" -> "1 sur 1" at a later beat) rendered the LATER value
baked in from frame 0 of an EARLIER, unrelated static-hold span, despite
the dedup log reporting "verified".
Root cause: computeStaticFrameSet's tween walker only tracks property
tweens, so a call()-driven textContent mutation carries no tracked
interval and the span around it looks fully static. verifyStaticFramesSafe
does catch genuine drift WITHIN a run it's checking, but a call() is a
one-shot side effect wired as both onComplete and onReverseComplete (GSAP
has no separate "undo" — crossing it in either direction fires the SAME
forward mutation). Verifying a LATER run forward-seeks past the call(),
permanently mutating the live page; an EARLIER run already passed its own
check before that happened, so nothing re-verifies it afterward. Real
capture then starts on the same corrupted page and bakes the wrong value
into the earlier span's reused buffer.
No reliable way to tell a DOM-mutating call() from a harmless one
(analytics ping, class toggle) without executing it, so this disqualifies
the whole comp on ANY call() — conservative, costs some dedup perf on
comps that use call() harmlessly, but correctness over speed.
Address two max-effort code-review findings on PR #2056 not covered by
the earlier review-gap commit:
- captureFrameToBufferPipelined's static-dedup reuse branch never
advanced session.lastEncodeResultFrame, unlike its sibling real-capture
branches. The gap-check window is computed from that watermark, so
every consecutive reuse in a static run rescanned an ever-widening
window instead of just the newest frame — O(n^2) total work over a
long static stretch instead of O(n).
- The "single-threaded, no race" justification on the shared
parallelGuard closure was wrong: the guard has real internal await
points (recapture, PSNR) between reading and writing its
sizes/absFloor/acceptedSmall state, so concurrent workers' calls do
interleave there (confirmed). Replaced with the actual reason it's
safe: absFloor only ratchets down, sizes is append-only and
order-independent for the median, and acceptedSmall's fast path
re-validates by exact byte-equality regardless of which worker wrote
the reference buffer.
Address PR #2056 review feedback:
- Fix totalFrames progress inflation for interleaved tasks — divide
each task's span by its frameStride to match the actual per-worker
frame count (captureFrameRange steps by stride), instead of summing
raw endFrame-startFrame which double(N)-counts interleaved tasks.
- Attach a no-op .catch to each frame's pipelined encodeResult at kick
time so an abandoned promise (loop exits early on abort/error before
draining it) can't surface as an unhandled rejection during teardown.
- Document why the pipelined branch's stride=1 path is validation-only
in production (HF_DE_PARALLEL_STREAM always uses interleaved
distribution) so a future refactor doesn't unknowingly widen it.
- Comment the intentional single shared parallelGuard/parallelStats
across workers (safe single-threaded, better rolling-median signal).
Step 2 of the DE engagement plan: multi-worker drawElement capture through
the streaming encoder, with the full runtime self-verification net riding
along — the confinement rule that kept the parallel clamp in place is now
satisfied on this path. Opt-in via HF_DE_PARALLEL_STREAM=true; default
routing (including the #2026 single-worker inversion) is unchanged.
Mechanism:
- distributeFramesInterleaved + WorkerTask.frameStride: worker i captures
frames i, i+N, i+2N... — seek-based capture makes stride free and the
ordered writer's reorder window shrinks from totalFrames/N to N (contiguous
chunks serialize workers behind the writer).
- Depth-2 pipelined worker-encode produce in the parallel worker loop (the
same shape as the sequential loop; frame k's in-page encode overlaps
k+stride's produce). HF_DE_PAR_DEBUG=1 traces the first frames per worker.
- Drain guard extracted to createDrainFrameGuard (session-parameterized):
every parallel frame gets the SAME blank-guard + PSNR self-verify as the
sequential drain, against its owning worker's pre-injection ground truth
(all sessions arm identical sample indices from
CaptureOptions.compositionDurationSeconds).
- FrameReorderBuffer.abort(err): a failed worker (e.g. verification error)
rejects all parked and future waiters — without this, peers park forever
in waitForFrame and the pool (which awaits ALL workers before surfacing
errors) deadlocks. Found by the verify-trip test; unit-tested.
- The typed DrawElementVerificationError is preserved past the pool's
error-string flattening so the orchestrator's verify-retry recognizes it.
- Static-dedup stride hazard fixed: lastEncodeResult reuse now requires EVERY
frame in (lastEncodeResultFrame, i] to be predicted-static (sequential
capture reduces to the old has(i) check).
- Workers get separate browser PROCESSES under the flag: pages co-tenant in
one browser starve non-active pages of BeginFrames on the paint-wait path
(measured 86s vs 30s on a 3,245-frame rAF comp).
Validation:
- Happy path W3: verify samples pass across workers (4x inf on the 2,381f
comp), output vs single-worker DE = 59.3dB (encode noise floor) — the
interleave + dedup-stride produce identical pixels.
- Verify-trip (marginal comp + HF_DE_VERIFY_MIN_DB=45): fails at frame 649
(32.2dB < 45), peers abort instead of deadlocking, whole render retries
via parallel screenshot, RENDER_OK in 42.6s.
- Canary suite 7/7 with the flag off (default paths untouched); producer
orchestrator tests 99/99; engine suite 909 passed (14 pre-existing main
failures, stash-A/B verified); reorder-buffer abort unit tests.
Perf note: capture-only parallel speedup measured 1.38x (W2) / 1.52x (W3)
over single-worker DE in the spike; end-to-end numbers on this machine are
currently noisy (separate-browser init overhead + bench load) — clean
benchmarks before any default routing change. The flag stays explicit
opt-in; promoting it into the router replaces the #2026 W=1 pin for the
same cohort.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Miguel Ángel