Recoil bounded probe journal, 2026-11-08

The new probe journal retains both uncached three-source reads and their six proof records in one immutable transaction. It awaits a bounded durable flush before the gate's final ownership and freshness checks. Production storage still does not meet the existing 250 ms fence. No Recoil qualification or input admission is claimed. No original, game, Resume, M, input or replay was run. Full27, parity, audio and FPS remain open.

Evidence R: scratch/runs/20261108-recoil-probe-journal-1753/. Starting main: e00ca4e8c (full SHA in starting-git-sha.txt). The predecessor is published f48707051, documented in recoil-source-transport-20261108.md. measured-journal-regression.json is the independently audited measurement, including every trial and refusal.

Implementation and fences

tools/native-source-retention.cjs adds a probe-scoped journal with a 32 MiB record bound, exactly two ordered read groups, and at most six response and six proof records. Admission serializes each complete envelope/proof immediately; later mutation cannot change it. Counting admission records does not repeatedly parse already retained large WASM responses. Before flushing, validation requires three distinct source lanes and matching guest/decoder/compositor proofs in each read, unique response sequences, ordered groups, unchanged identities across reads, successful uncached responses, original timestamp brackets, complete source sizes and SHA-256 matches. Refusals and partial evidence are preserved in an explicitly incomplete diagnostic transaction; they cannot qualify.

The new file is exclusively created, written, fsynced and closed; its directory is then fsynced and closed so a durable file does not depend on an unpersisted name. This is one transaction flush, including the necessary directory barrier. Its <=250 ms monotonic bound includes validation/serialization. Timeout is a terminal refusal, even if a delayed timer callback loses a race to IO completion. The writer is awaited to settlement after failure or timeout. A timed-out flush can therefore take longer to drain than its admission deadline. It never returns a late certificate. A complete:true header means content completeness only; it is neither a successful durability receipt nor a certificate. There is no journal recovery path that can certify or replay an abandoned probe.

native-software-gate.cjs uses the journal for responseDirectory probes. Pre-probe source discovery and explicit legacy callback callers retain their existing individual sinks. strict-software-gate.cjs opens the transaction, labels its two reads, flushes after actual PNG decode and all existing evidence saves, then runs the existing current page/software/guest owner, instance, script/context, retirement and freshness checks. Both reads' original request ages, both copy ages and PNG age through decode/save remain <=250 ms. No timestamp is refreshed for durability. The normal independent challenged clock wrapper still runs its fresh post-probe census, rechecks all these ages and enforces the immutable five-second compound window. Neither clock producer nor kernel/parent custody was changed. Existing source lanes use allSettled; failure additionally awaits the diagnostic journal/writer drain, closes the gate and forbids replay.

Actual Chrome measurement and remaining cause

One fresh box ran two finite controlled fixtures under the existing kernel custodian, production parent census/subreaper and runtime lock. Each alternated journal/legacy/journal/legacy using actual Chrome Debugger.getScriptSource calls. The exact accepted WASM (2,182,663 bytes, 6cf120d0) and decoder/compositor (f468ffce / cb21daaf) were compiled/read, without executing a guest. Their existing retained module/source paths were reused, independently hash-checked. The causal VM supplies synthetic owner/instance objects and a decoded 1-pixel PNG; the normal native adapter, strict Software gate, phase-save callback and independent clock guard run their real code. This measures their complete source/evidence path, not actual game ownership, HUD or Software qualification.

All 36 complete envelopes and 36 source proofs were recovered and audited. Every journal trial has six envelopes and six proofs with their original request/response timestamps and full accepted byte hashes. Legacy trials stopped after the first three-source read when their existing gate expired; they were not forced into a six-response success. All eight trials refused certification. Remote-a preceded tightening the flush bound to include preparation and an explicit elapsed-time check; remote-b exercises those final semantics. Small intermediate patches reconstruct both measured retention versions and the fixture instrumentation exactly, verified against their actual provider hashes. The final source differs from remote-b only in its explanatory header comment.

Journal trial Two-read source window Journal IO, excluding preparation/open Outcome
remote-a / 1 219 ms 102.33 ms source freshness expired
remote-a / 3 240 ms 82.33 ms source freshness expired
remote-b / 1 276 ms 224.08 ms bounded flush timeout; writer drained
remote-b / 3 226 ms 180.49 ms source freshness expired

Source-message intervals in the four journal trials were 72.57–110.99 ms per lane. They do not include every later retention/proof/save step. File fsync alone was 53.52, 53.66, 156.17 and 137.70 ms; directory fsync was respectively 31.57, 16.03, 54.08 and 32.10 ms. These actual durable waits forbid a latency success claim from batching alone.

The instrumented remote-b trials identify the additional normal-driver cost. After the final source response, durable completion took 469 / 523 ms; the oldest original source was then 745 / 749 ms old. The remaining owner-save, PNG/pixel-save and return-save barriers consumed 82.45+66.21+65.89 ms and 88.39+127.98+103.14 ms, respectively, before journal validation/write/durability. The save between source reads cost 56.03 / 45.60 ms, increasing first-to-second request separation to 170 / 138 ms. One trial was already beyond 250 ms at its last source response, before the final journal. Entry-save costs of 83.94 / 108.71 ms are outside source age but still inside the compound window. This is measured evidence of remaining per-phase durability plus journal storage service time, not a conjectured Chrome-only transport delay.

The implementation-ready next correction is to stage the existing phase JSON and the actually decoded PNG as immutable records in the same transaction, while preserving every original timestamp and phase payload. Replace their individual fsyncs with that transaction's final barrier. Pre-create and durably name a bounded append journal before the compound window, then stream immutable frames asynchronously while source reads proceed; validate lengths/hashes and await one final data sync/close before the unchanged current-owner/census/freshness gates. Frame bounds, aborted/partial writes, crashes and writer drain need the same negative controls. This removes the measured intermediate phase barriers and per-probe directory sync; it must be measured again. Even that cannot be promised to meet 250 ms on this volume: the observed 137–156 ms file sync leaves only 94–113 ms for both fresh reads, remaining checks and independent census. If overlap/preallocation cannot reduce that measured service time, a genuinely faster durable volume is required. A tmpfs substitution or relaxed freshness is not a correction.

Selection, tests and lifecycle

selected-runtime-pins.json starts with the assigned selected1067 manifest. Only the native adapter, retention module and strict gate change among shared providers. The exact prior next-driver-retention.patch was applied to the new physical driver only; its 29-helper closure was normally relocated and freshly pinned. All 1,067 providers resolve and hash-check; normal physical closure passes. Original851, accepted1066 (position-runtime-pins.json, not the nonexistent runtime-pins.json) and selected1067 predecessor manifests remain immutable. No new original was started. relocated-helper-changes.patch records the small helper changes; the remaining physical helpers are the normal runtime closure, not a source-tree archive.

Focused checks pass: eight journal/sink test entries, including actual process crash before fsync, exclusive refusal after crash, immutable complete groups, partial/duplicate/mismatched/refused records, byte bounds, write/fsync/close and directory failures, timeout writer drain, failed source lanes and retirement through flush. Native13, strict46 and owner28 negative controls pass, preserving actual PNG decoding, allSettled, sticky UNKNOWN and no replay. Clock, runtime selection and physical helper closure tests pass. The final locked full build passes; logs and exact relevant binary diff are in R.

Initial local fixture attempts are retained: local-a stopped on a fixture copy-before-install error; local-b used the fixture's tiny module pin and was refused. Both drained normally. local-c supplied the corrected accepted pins; its results preceded the removal of repeated admission parsing. These are not accepted production performance comparisons. Final local-d uses the final implementation: both journal probes refused freshness, while both legacy baseline probes returned unqualified results after their fresh final censuses. Local logs remain separate from the production measurements. The first clock test invocation omitted its required environment and stopped before Chrome; the corrected locked invocation passed all actual Chrome clock controls.

Both remote kernel drains exited 0: 158/158 and 176/176 owned/reaped tasks, empty remaining/unresolved lists. This does not audit CLONE_UNTRACED or establish exhaustive lineage outside the custodian's stated scope. No outer runtime timeout killed either supervisor. Before stop, independent checks found 338 recorded PID/start identities absent, all four ports closed, both profiles absent, no runtime FD writers and the runtime lock free. All 136 unique remote outputs / 89,762,902 bytes were retrieved and independently reopened/hash-verified before the delegated normal stop.

Box bx_q54tv2w3 is ARCHIVED, current expiry NULL; initial expiry was 2026-11-08T18:24:09.226Z. Normal stop stop_0f0339ee0901f6c3 completed 18:04:26.588Z, forced FALSE, error NULL. The observable independent 25-minute watchdog preceded the one persisted UUID-idempotent create; actual ID/expiry preceded runtime. Local lifecycle roles exited 0 and their supervisor recorded no remaining children. Exact Git SHA, relevant binary diff, small owned patches and existing source/module path/hash references provide recovery without source archives, media duplication or peer WIP. No old box resume/delete, deployment, private game writes, clock/seed change or subworkers were used.