Recoil uncached source transport, 2026-11-08
The production driver synchronously fsyncs complete source envelopes and source proof records on its CDP event loop. Those disk waits delay delivery of other already requested source responses. A controlled remote Chrome comparison isolated and repaired this transport contribution. Actual Recoil Software qualification remains open: durable evidence completion still exceeds 250 ms. No original, Resume, M, gameplay input or replay was launched for this task.
Evidence R: scratch/runs/20261108-recoil-source-transport-1728/.
Starting Git SHA: f123e073f0ecca088f581c61c9b12cedb3ed8664.
The actual preceding generation is the immutable 1,066-provider manifest at
20261108-recoil-original-1702/position-runtime-pins.json (SHA-256
a520a909a0df912875d130a50a70085c559fa57aa7cde7095c0af161782656a8).
The task's runtime-pins.json spelling did not exist; reference-pins.json
records the physical manifest and exact relevant rows. Historical original851,
kernel2486507db, producer68ea2f56 and endpoint11683651 remain unchanged.
The original failure's 442 ms is a driver-observation interval, not a measurement
of Chrome alone. In browser-position.cjs, WebSocket onmessage parses JSON and
resolves a pending command. The native Software adapter then clones and retains
the complete response before returning it. Both the source retention callback
and the subsequent proof-record callback use writeSync/fsyncSync. While either
callback waits on disk, the same event loop cannot receive the other CDP lanes.
The response envelope's responseAt precedes that envelope's own retention, but
can include earlier lanes' synchronous retention and proof writes.
The first local Chrome fixture used the actual kernel and parent custody code. It did not reproduce production storage latency: concurrent reads completed in 87–129 ms through retention/hash; switching native WebSocket to ws had no consistent benefit. No WebSocket replacement was made. The initial local fixture used the checkout's newer JS; it is only a preliminary control. The remote comparisons use the exact accepted JS recovered from complete original responses57/58, independently hash-checked: decoder204938/f468ffce and compositor227919/cb21daaf; WASM2182663/6cf120d0 is read from the existing retained module path. No game media or source-tree archive was copied.
One newly leased box ran finite transport fixtures under the unchanged ptrace kernel and production parent census/subreaper code. The parent wrapper adapts only the nested fixture directory and uses the normal runtime lock. The independent persistent challenged clock producer starts before Chrome. Every source response comes from a fresh actual Debugger.getScriptSource request; allSettled drains each three-lane group. Exact source strings were compiled in an isolated fixture page and the real WASM module was compiled without running it. This exercises CDP transport and evidence IO, not actual guest instances, Software pipeline owners, or game qualification.
The final remote-d comparison alternates synchronous and asynchronous sinks
twice in the same Chrome, box, storage and custody path. Each round has one
guest-only control followed by two complete three-source reads. All28 complete
response envelopes and all28 proof rows are retained and independently audited.
The decisive measured results are in causal-transport-regression.json:
| Stage | Synchronous sinks | Asynchronous sinks |
|---|---|---|
| Concurrent WASM request to message handler, four reads | 162–238 ms; median212 | 92–102 ms; median99 |
| Source-envelope fsync | 28–84 ms | 49–185 ms, off event loop |
| Proof-record fsync | 16–63 ms | 21–114 ms, off event loop |
| All source/proof durable completions | 58–346 ms | 162–369 ms |
JSON parsing across final source responses was 0.21–1.86 ms; envelope stringify was at most5.75 ms and proof stringify at most0.05 ms. Parent census cycles in the source interval were55.3–58.9 ms, median57.4, including the fixed50 ms sleep; 31 process identities were in those parent snapshots. Three kernel events were observed in the measured interval and are retained verbatim. These observations do not establish exhaustive kernel lineage or separate every Chrome scheduling cost. The transport repair is supported by the alternating comparison and measured synchronous disk stalls, without removing either custodian.
tools/native-source-retention.cjs provides asynchronous complete envelope
write/fsync/close and an ordered asynchronous proof journal. Both are awaited;
exclusive source creation prevents overwrite, and a proof-journal failure
poisons its queue. native-software-gate.cjs accepts explicit
responseDirectory to select these sinks. Existing callback callers continue
to work. The accepted original driver is deliberately immutable; the exact
two-callback replacement is supplied as next-driver-retention.patch for the
next normal relocated generation, not silently applied to the old manifest.
selected-helper-runtime-pins.json explicitly selects the changed native adapter
and adds the new sink:1,067 providers. The original851 and every other existing
row are unchanged. All1,067 providers were resolved and hash-checked, and the
ordinary physical helper closure passed. This is a preparatory helper selection;
the next relocated driver must apply the supplied patch and be freshly pinned
before launch. No original was launched with these helpers.
Focused tests passed: six Node test entries, including complete two-read/six response retention, Unicode and full WASM-sized envelopes, exclusive creation, durable failure propagation, ordered proof appends and poisoned-queue refusal. Existing native13, strict Software46 and owner28 negative controls passed, including allSettled, context/instance retirement, delayed sinks, PNG decoding, sticky UNKNOWN and no replay. The final locked full build passed. The six uncached responses, original proof timestamps, independently fresh challenged census, actual owner checks, final PNG through decode/save <=250 ms and immutable five-second compound window are unchanged.
This fixes event-loop transport interference, not the storage service time. The asynchronous durable completion range itself forbids a qualification claim. The concrete next implementation is a probe-scoped durable journal transaction: retain immutable complete envelopes and proof rows for both fresh reads, then perform one bounded durable flush before the certificate can return or any input can be admitted. It must retain every original request/response timestamp, await all failed lanes and writer drain, reject incomplete six-response groups, and run the existing final ownership/retirement/freshness checks after the flush. Failure or timeout must remain sticky UNKNOWN. That changes the current per-response durability barrier and therefore needs its own fail-closed gate tests before use; it was not smuggled into this transport-only patch. No claim is made that batching will meet250 ms without measuring it.
Operational failures are retained. Remote-a's90-second runtime/110-second outer fixture cap killed its kernel supervisor during shutdown; EXITKILL retired its tracees. It is not a normal kernel drain. Independent checks found all158 recorded task identities absent before the next fixture. Its leftover owned Chrome profile was removed only after independent PID/start, endpoint and FD writer checks; the initial profile-absence refusal is retained. Remote-b stopped on a fixture-only sequence filename collision; its partial timings are not the accepted regression. That collision was corrected before complete remote-c/d. Remote-b/c/d kernel drains are exit0, respectively154/156/156 owned tasks reaped. Remote-c retains two unresolved autoattach lineages; final remote-d is lineage complete. CLONE_UNTRACED remains unaudited. Outer launcher start ticks were not recorded; their PID-only disappearance is not represented as an identity proof.
All249 unique remote outputs /188,746,303 bytes were retrieved and independently reopened/hash-verified before normal API stop. Final independent census checked 628 task/supervisor PID/start identities absent; eight ports closed, all owned profiles absent, no runtime FD writers, runtime lock free, no input sent. Box bx_2qvsxyyn is ARCHIVED, expiryNULL; initial expiry 2026-11-08T17:58:12.388Z. Normal stop stop_6d70e1f754f84c70 completed 17:43:36.900Z, forcedFALSE/errorNULL. The observable25-minute lifecycle was prearmed before the persisted UUID-idempotent create; ID/expiry preceded runtime. All local lifecycle roles exited0 and the supervisor recorded no children. Exact Git SHA, relevant binary diff, small owned patches and existing binary path/hash references are retained without source archives or repeated accepted logs. No old box resume/delete, deployment, private guest writes, clock/seed change or subworkers were used. Full27, parity, audio and FPS remain open.