The graph rides 0.5.1: polyengine#239 closes the checkpoint's starvation of its sibling drivers - #104
Merged
Merged
Conversation
…ion of its sibling drivers The field report was 'driveAsync: a resumed-activation claim was never released' on mobile Firefox during Drive testing, and both halves of the obvious reading were wrong: not Firefox (a nine-case JSPI ordering probe shows SpiderMonkey and V8 byte-identical on every primitive this rests on), and not network latency (a first diagnosis read its own 4s proxy as the cause; the control run fires the assert at ZERO delay). The mechanism, proven causally on stock 0.4.0: driveAsync's unconditional speculative pending-resumption entry is a store-wide gate. The device's own two periodic drivers collide with it — the worker's 500ms-debounced NON-BLOCKING checkpoint (a recorded design decision) is near-always in flight, and the solo page's 1s us-events drain arrives into it and can only hop; ten thousand hops later the assert fires with the user's exact message. A phone is simply where slow OPFS widens the window. Deleting the one speculative-entry line takes the reproduction from 18 firings to 0; upstream #239 (in v0.5.1) guards exactly that line and adds the driver-arrival wake. Matrix row 47b is the regression gate and measures the starvation itself: 12s of back-to-back checkpoint + us-events on one store. On 0.4.0: one drain in twelve seconds, then the trap. On 0.5.1: 108 drains, no trap — the drain count recovering is specifically the driver-arrival wake, the half a delete-the-line test cannot exercise. The bump rode A22 (embedder is application-only now): the error vocabulary — ComponentException, isComponentException, isTrap, toCloneable, fromCloneable — moved to @polyengine/protocol across eleven files, and the pin-comment's module-identity reasoning was rewritten for the new shape (ports couple to the protocol package, not the runtime; the one-runtime obligation stands on its own). The webrtc sibling advances to v0.5.0 — the only resolution that keeps one runtime in the graph AND the lazy-bundling property (JSR's 0.5.0 re-trips the documented node: bundle hazard; the CI sibling pin follows). wasi/fs's in-window mutate-directory tightening lands on the permissive side of our writable:true preopens — checked, not assumed, since the previous bump's checklist wrongly claimed an empty window. Harness, from the same investigation: Playwright Firefox gains the persistent-storage pref (navigator.storage.persist() never settles headless without it; four more scenarios now pass under Gecko) and a recorded Juggler hazard (promising exports called from page.evaluate frames SIGSEGV the content process — Firefox-lane scenarios drive through page scripts). Gates: devstore matrix ALL PASS with 47b red-then-green across the bump, e2e 24/24, all four bringups, engine check/resume/pair, invariants 9/9, resolved graph exactly one runtime@0.5.1 and one protocol@0.2.3.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes the field-reported crash (mobile Firefox, Drive testing):
driveAsync: a resumed-activation claim was never released.Both halves of the obvious reading were wrong, and the round's record says so: not Gecko (nine-case JSPI ordering probe: SpiderMonkey ≡ V8), and not network latency (the first diagnosis read its own 4s proxy as causal; the control fires at zero delay — retracted). The mechanism, proven causally on stock 0.4.0:
driveAsync's unconditional speculative pending-resumption entry is a store-wide gate, and the device's own two periodic drivers collide with it — the worker's deliberately non-blocking 500ms checkpoint is near-always in flight, the solo page's 1sus-eventsdrain arrives into it and can only hop, and 10k hops later the assert fires. A phone is just where slow OPFS widens the window. Upstream polyengine#239 (in v0.5.1) guards exactly the proven line and adds the driver-arrival wake.Row 47b measures the starvation itself: checkpoint+us-events storm, 12s. On 0.4.0: 1 drain, then the user's exact trap. On 0.5.1: 108 drains, no trap — the drain recovery specifically exercises the wake, the half a delete-the-line causal test can't reach.
What rode along:
@polyengine/protocol(new pin, 11 files). Pin-comment module-identity reasoning rewritten with was-true-until history (ports couple to protocol now; the one-runtime obligation stands on its own).node:bundle hazard). CI sibling pin updated; local sibling checkouts must advance to v0.5.0.mutate-directorytightening: checked against ourwritable: truepreopens — permissive side, no change; recorded because the previous bump's checklist wrongly claimed an empty wasi window.persist()never settles headless without it; solo-persistence/storage/gdrive/account-storage all pass under Gecko now) and a recorded Juggler hazard (promising exports frompage.evaluateframes SIGSEGV the content process).Gates: matrix ALL PASS (47b red→green across the bump), e2e 24/24, four bringups, engine check/resume/pair, invariants 9/9, resolved graph = exactly one
runtime@0.5.1+ oneprotocol@0.2.3.