Recoil durable extent preallocation, 2026-11-08
Real extent preallocation is implemented and tested. The complete unchanged 250 ms fence still refuses every measured trial. The fresh-box preallocated trials reached durable close at source ages 317 and 370 ms; final fsync alone cost 122.56 and 130.39 ms. The fastest compound baseline closed at 233 ms, then its independent post-probe census correctly refused at 260 ms. No original, game, Resume, M, input or replay was run. Full27, parity, audio and FPS remain open.
Evidence R: scratch/runs/20261108-recoil-preallocated-journal-1844/.
Starting Git SHA: f37f9c8827d3981a4c024a4e304d2c5be11b0661.
The exact published compound baseline is e71fbddbe, described in
the compound journal note.
measured-preallocation.json, measurement-audit.log and the retained complete
journals provide the measured results and independently checked source hashes.
Supported helper and durability ordering
tools/native-source-retention.cjs uses an explicit supported normal helper:
Linux, /usr/bin/python3 -I, and Python's os.posix_fallocate. The already
exclusively opened journal file description is inherited as FD 3. No shell or
pathname reopen occurs. Missing interpreter/API, unsupported filesystem,
allocation error, abnormal helper exit or timeout refuses preparation. There
is no sparse-truncate or zero-fill replacement. A five-second preparation cap
kills an expired helper and waits for its close before file cleanup. The helper
exits before the existing file/name syncs and before compound timing. Its
stderr is bounded, and size plus allocated-block checks reject an unallocated
file. The normal adapter's existing error path makes refusal sticky UNKNOWN,
drains the writer, and prevents replay.
The default bound remains 32 MiB. Before the guarded compound window, preparation reserves that file's real extents, syncs the file, and syncs/closes its directory. The normal driver retains its existing prepare/abort callbacks. Immutable frames stream by explicit positions captured at admission. A short write advances the buffer offset and file position only by the returned count; zero/invalid progress refuses. After the terminal seal and all queued writes, finalization truncates the file to its exact sealed byte length, then performs the same final fsync and close. All final owner, context, instance, script, retirement, independent census, source/PNG <=250 ms and immutable five-second checks remain downstream. Original response/copy timestamps and actual PNG decode/save are unchanged.
A crash before exact truncation leaves an exclusive diagnostic file, possibly with a reserved zero tail; the diagnostic reader refuses partial/corrupt metadata. No abandoned-file recovery can certify or replay. Error and timeout paths still wait for queued writers and close. A complete seal describes content, never qualification or successful durability by itself.
allocation-syscall.* independently records the real successful syscall
fallocate(3, 0, 0, 33554432) = 0, positioned writes, exact ftruncate, fsync and
close. The tested offset is deliberately advanced before admission to check that
frame writes ignore the shared seek position. retrieved/external-helper.json
records the remote interpreter path, resolved binary, SHA-256, size and version:
/usr/bin/python3.12, SHA-256
1643dacd9feaedc58f3cc581e4d22577dfe25c09b10282936186ccf0f2e61118.
The supported mechanism is intentionally fail closed outside this environment.
Finite actual Chrome comparison
One fresh box executed exactly two alternating preallocated/compound pairs under the same published kernel, production parent, normal native adapter/strict gate, actual Chrome source transport and independent challenged clock producer. The baseline reconstructs the exact published compound modules in memory from current providers plus small verified reverse edits. The preallocated module uses its normal default helper; neither trial changes fences or evidence callbacks. Preparation costs 1526.88/678.38 ms for the preallocated trials, outside the compound clock, versus 116.60/265.54 ms for the baseline. Preparation includes normal kernel tracing and durable file/name work; it is not an allocation-only microbenchmark.
| Fresh-box trial | Two-read source window | Exact truncate | Final fsync | Oldest source at durable close | Post-census age |
|---|---|---|---|---|---|
| Preallocated 1 | 183 ms | 0.085 ms | 122.56 ms | 317 ms | refused before census |
| Compound 2 | 181 ms | unchanged baseline | 43.22 ms | 233 ms | 260 ms, refused |
| Preallocated 3 | 227 ms | 0.053 ms | 130.39 ms | 370 ms | refused before census |
| Compound 4 | 195 ms | unchanged baseline | 118.87 ms | 327 ms | refused before census |
Every trial retained six complete uncached source envelopes, six proofs, the five
phase payloads and the actual decoded PNG. Independent audit checks original CDP
timestamp brackets, byte counts and all accepted hashes: WASM 6cf120d0, decoder
f468ffce, compositor cb21daaf. The 2,182,663-byte accepted WASM is reused by
existing path/hash; the current full-build module is not substituted.
The real Chrome/source/storage path is measured; ownership and the one-pixel PNG
belong to the causal VM fixture. These are not original-game ownership/HUD proofs.
A separate local pair comparison also refused all four trials. Preallocated source windows were 241/246 ms, final fsyncs 15.49/12.71 ms and durable-close ages 271/270 ms. Baseline windows were 292/231 ms, syncs 18.55/13.00 ms and ages 321/254 ms. Local uses the current shared kernel and is kept distinct from the fresh-box comparison's exact historical kernel. Across both runs, all 48 complete envelopes and 48 proofs, eight complete journals and eight decoded PNGs were independently audited. No trial passed the complete fence.
Preallocation removes extent reservation from the timed window, but these finite samples show no reliable end-to-end improvement. The fresh-box final durable barrier remains 122–130 ms despite successful preallocation; truncate costs <0.1 ms. This does not isolate filesystem internal causes or prove a particular volume will be faster. The baseline's 233-to-260 ms interval demonstrates that the independent census remains necessary even when the strict gate first passes.
Remaining service need and safe next step
For preallocated trial 1, the measured budget is 183 ms through the final source, about 11.4 ms other work, and 122.56 ms fsync. Adding the observed 26.53 ms independent census would require final sync below about 29 ms merely to fit 250 ms, before a safety margin. Trial 3 already uses 227+12.6+26.5 >250 ms without any sync cost; storage alone cannot rescue that source window. Original/copy/PNG ages and other checks can independently be tighter. This is a measured budget, not a qualification guarantee.
The useful safe next step is a separately assigned comparison on an actually
different durable low-latency block service or improved durable I/O service
quality. Measure the same accepted Chrome sources, helper, kernel, final
fsync+close and independent census, targeting <=10 ms final fsync and consistently
short enough source windows (roughly <=200 ms with the observed remaining work;
less for usable margin). Retain producer-proof durability and every present
fence. This box had no separate volume measured; predecessor volume inspection
found /home and /tmp on the same device. A renamed same-device path, tmpfs,
restamping, omitted census, relaxed source fence or repeated original cannot
supply the missing durable service. No original attempt is warranted here.
Validation, selection and recovery
Thirteen focused retention tests pass, including actual partial-frame process crash, exclusive abandoned-file refusal, immutable frames, byte/row/time bounds, explicit short/zero writes, open/allocation/write/truncate/sync/close/directory failures, real readonly/closed-FD syscall refusal, helper timeout with child-close drain, late writer/close drain, blocked timer races, prepared-census failure, retirement/context changes and source/PNG/five-second expiry. Native13, strict46, owner28 and lifecycle12+6 controls pass. Actual Chrome clock controls pass with seven independent process proofs and the final physical driver binding. Physical closure13, generation selection8 and the locked full build pass.
The full-suite strace invocation correctly hit the unchanged 250 ms timeout
under thousands of instrumented short writes. The targeted real-allocation trace
passes; the ordinary complete suite passes. Initial selection and standalone
watchdog test invocations lacked their required import/argument/environment
configuration; those failures are retained without being counted as passes.
Unchanged watchdog/clock/gate proofs are referenced and hash-checked in
unchanged-proofs.json; no redundant watchdog runtime was needed. The task's
allocation watchdog and measured runtime's watchdog were independently prearmed
and observable, and actual shutdown completed normally.
selected-runtime-pins.json and position-runtime-pins.json retain 1,067 providers
and relocate the normal 29-file physical helper closure. Only retention code
changes among shared providers. The two-callback and prepare/drain driver patches
are byte-identical to the predecessor. Original851, accepted1066 and historical
selected1067 remain immutable. The shared kernel changed during other work;
remote provision uses its exact e71fbddbe Git bytes, matching the predecessor's
150822f3 hash, and leaves the shared newer file untouched. Every selected
provider was verified, using the recorded historical Git fallback when necessary.
Unchanged proofs are reused; no fixture churn or peer WIP is included.
Git holds the source archive. starting-git-sha.txt, evidence-git-sha.txt,
relevant-uncommitted.patch, baseline reverse edits, helper relocation patch and
small run scripts record exact recovery. Existing module/source paths and hashes
are retained rather than copied into evidence. There are no source-tree/media
archives, old-run deletion, box resume/delete, public deployment or subworkers.
The run remained above 1.5 GB free on /home, with no RAM gate.
Retrieval and delegated shutdown
All 53 remote outputs / 27,707,189 transported bytes were retrieved and
independently reopened/hash-verified before normal stop. The unique 25,249,889-byte
custody history is individually gzip-retained. Its original SHA-256 is
0558adffff28cd7922867d22e223bcb754798e89baae247b70e622b6410303b1;
gzip-originals.json records the recovery recipe. Independent streaming
decompression verified original size/hash without a durable uncompressed copy.
The final audit rereads every retained task artifact and every unique remote
output, including that decompressed SHA.
Kernel drain was 179/179, exit 0, remaining empty and unresolved lineage empty. Independent remote checks found 181 observed PID/start identities absent, two ports closed, the Chrome profile absent, no runtime FD writers, no held input and the runtime lock free. CLONE_UNTRACED is unaudited: this is not an exhaustive process-history claim. Independent local drain checks 166 observed identities absent, profiles/ports/writers drained, lock free and the allocation supervisor empty. No board watcher was left running.
Box bx_gbee99be is ARCHIVED, current expiry NULL, initial expiry 2026-11-08T19:10:12.489Z. Delegated normal stop stop_1569916c99242c5e completed 18:47:22.539Z, forced FALSE, error NULL. The independent 25-minute immutable watchdog was observable before the persisted UUID allocation key and creation. Actual ID/expiry were retained before runtime. The worker releases this task and stops; the selected generation remains unqualified for gameplay input.