Recoil compound evidence journal, 2026-11-08

The compound journal removes the measured individual phase-save barriers and moves durable name creation before the guarded compound window. Qualification still fails the unchanged 250 ms fence. The fastest final trial reached durable close at source age 230 ms, then the independent census correctly refused at 268 ms. No original, game, Resume, M, input or replay was run. Full27, parity, audio and FPS remain open.

Evidence R: scratch/runs/20261108-recoil-compound-journal-1823/. Starting Git SHA is in starting-git-sha.txt; predecessor is published 706a23082, described in recoil-probe-journal-20261108.md. measured-compound-regression.json contains the independently audited trials.

Implementation and unchanged fences

native-source-retention.cjs now streams an exclusively created, bounded 32 MiB version-2 framed journal. Before the compound clock starts, prepareProbe() creates its file, syncs it, syncs/closes its directory and keeps the exclusive file descriptor. Each admitted frame owns immutable bytes and enters an ordered asynchronous write queue. Source callbacks and phase saves do not await storage. Short writes advance only by the reported count; zero/invalid progress refuses.

One complete transaction holds six complete uncached CDP response envelopes, six matching proofs, the five original phase JSON payloads (entry, owner/read1, owner/read2, pixels, return), and the actual PNG bytes passed to the sink after real PNG inspection. The phase payloads retain their original content, including the embedded pixel data URL. The separate binary PNG must match its inspected result hash. Frame lengths, payload hashes, record counts and a terminal whole-transaction hash make partial/truncated/corrupt diagnostic evidence recognizable. There is no abandoned-journal certification or replay path.

Admission retains the original request/response timestamps; validation still checks both complete three-lane groups, identities, hashes, byte counts, ordered unique sequences, original timestamp brackets and uncached success. Final validation, queued writes, terminal frame, file sync and close all precede the strict gate's existing final owner/context/instance/script/retirement checks. The <=250 ms monotonic flush deadline includes validation and preparation of the seal; explicit elapsed checking also refuses delayed timer callbacks. Timeout/error awaits writer settlement and close before leaving terminal UNKNOWN. A complete seal describes content, never durability success or qualification.

The native adapter replaces the driver's individual phase-file save callback with this journal sink in responseDirectory mode. Explicit legacy callbacks remain available outside that mode. Source lanes retain allSettled drain. Preparation/census failure also poisons and drains an otherwise unused prepared journal; it cannot leave an open writer before a later replay. The new physical driver prepares before calling the guarded probe and drains on guard failure.

The strict gate, clock producer/verifier, kernel custodian and production parent are unchanged. Both original source reads, both copy ages and PNG freshness through actual decode/save still face the 250 ms checks. The independent challenged producer still performs its fresh post-probe census and durable proof retention, rechecks source/PNG ages, and enforces its original immutable five second transition. No restamping, cache, tmpfs, alternate clock, relaxed fence or retirement exception was introduced.

Actual Chrome comparison

One fresh box ran two finite alternating compound/published-journal comparisons under the normal kernel and parent custody harness. Both use actual Chrome Debugger.getScriptSource for the accepted 2,182,663-byte WASM (6cf120d0) and actual decoder/compositor sources (f468ffce / cb21daaf). Existing module/source paths and hashes are retained without repeated media. The normal native adapter, strict Software gate, actual PNG decoder, phase payloads and independent clock guard execute their real code. Ownership/instance objects and the one-pixel PNG belong to the causal VM fixture, not an original Recoil run. This measures the normal helper's evidence path; it does not qualify actual game ownership or HUD.

All eight trials retained six complete envelopes and six proofs: 48 of each were independently audited against accepted byte hashes and original uncached CDP brackets. All four compound journals independently decoded to five phases and the actual CRC-checked PNG. Every trial refused qualification.

Run / compound trial Two-read source window Final file sync Oldest source at durable close Outcome
remote-a / 1 228 ms 210.94 ms 455 ms strict freshness refusal
remote-a / 3 174 ms 99.28 ms 286 ms strict freshness refusal
remote-b / 1 176 ms 41.58 ms 230 ms post-census refusal at 268 ms
remote-b / 3 179 ms 87.04 ms 281 ms strict freshness refusal

Remote-b uses the final cleanup semantics. Remote-a differs only in prepared journal cleanup and the baseline fixture loader; exact provider pins and a small verified recovery patch reconstruct it. The baseline is the exact published 706a23082 journal/native code with unchanged strict gate. Final baseline modules are reconstructed in memory from current paths plus verified small reverse edits.

In remote-b, published-journal source windows were 226/235 ms and oldest sources at durable completion were 450/488 ms. Its final owner/pixels/return saves cost 23.02+15.22+45.98 = 84.22 ms and 24.74+52.98+73.18 = 150.91 ms; the between-read save cost 55.07/69.33 ms. Those per-phase fsyncs are absent from both compound trials. Preparation (85.28/83.02 ms) precedes the guarded window. The compound final-source-to-durable-close intervals are now 54/102 ms, versus 224/253 ms for the baseline. This removes the demonstrated phase-save bottleneck without claiming a stable wall-time speedup from four variable-storage pairs.

The exact remaining fast-trial budget is 176 ms through the final source, about 12 ms of staging/validation/writer completion, 41.58 ms final sync, then about 38 ms for the independent final census/retention and checks: 268 ms. The second trial's 87.04 ms final sync already expires before that census. This identifies both the remaining storage service time and the required post-census margin; bypassing the census would be an invalid correction.

Durable storage alternatives and safe next correction

retrieved/durable-volume-measurement.json records 24 finite storage-only trials with 7 MiB of random synthetic data. Both /home and /tmp are durable ext4 on the same /dev/vda1 device (64769); /tmp is not substituted into the gate. findmnt/lsblk show no separate usable evidence volume attached. The EFI boot partition is not a candidate evidence volume; tmpfs mounts were excluded.

For /home, ordinary final fsyncs measured 22.44–48.41 ms, fdatasync 20.48–39.54 ms, preallocated fsync 19.55–23.02 ms, and preallocated fdatasync 19.81–27.69 ms. /tmp ordinary fsync measured 19.90–34.98 ms and preallocated fsync 22.65–26.50 ms. These serial storage-only samples cannot replace the actual Chrome syncs of 41.58–210.94 ms, or establish a different physical volume's speed. No physically different usable durable volume was available to compare.

A concrete safe next code correction is to reserve the bounded file's real extents with posix_fallocate before compound timing, sync its file/name there, stream by explicit offsets, then truncate to the exact sealed length and perform the same final fsync+close before all current checks. It must keep the 32 MiB bound, crash/partial/error controls and immutable source timestamps. Storage-only preallocation improved some samples but remains unqualified with Chrome/census; this change is not implemented or represented as a gameplay-ready generation. For the measured 176 ms source window, a durable service with <=10 ms final sync would leave roughly 24 ms margin after the observed 38 ms census and 12 ms other work; that is a measurement target, not a guarantee. If preallocation cannot meet the complete fence, attach and measure a separate low-latency durable block volume with those exact checks. Renaming a same-device path, using tmpfs or removing producer-proof retention does not address the remaining cause safely.

Selection, validation and lifecycle

selected-runtime-pins.json and its normal position-runtime-pins.json alias contain 1,067 verified providers and 29 relocated physical helpers. Only the native adapter and retention module change among shared providers. The new physical driver preserves the prior two-callback patch, adds preparation and failure drain, and is freshly pinned with its changed closure. Original851, accepted1066 and the selected1067 predecessor remain immutable. Existing unchanged source/kernel/clock proofs and accepted module pins were reused. The current full-build WASM is not substituted for the selected accepted WASM.

Nine focused journal/sink tests pass: actual process crash during a partial frame; exclusive abandoned-file refusal; immutable phases/PNG/sources; partial, corrupt and oversized frames; byte/row/time bounds; short/zero writes; open, write, sync, close and directory errors; late write/close completion; blocked sync timer races; prepared-census failure; and retirement/context/age faults through final flush. Native13, strict46, owner28 and lifecycle12+6 controls pass. Actual Chrome clock controls pass with seven retained independent process proofs and the final physical driver binding. Runtime selection8, physical closure13 and the final locked full build pass. The first clock invocation lacked its required driver-path environment; its failure and normal teardown are retained separately. The corrected final invocation passes.

The local comparison refused all four probes (one compound trial expired after its first read); it remains distinct from the durable-box comparison. Both remote runs completed normal kernel drain: 178/178 and 159/159, exit 0, remaining empty and unresolved lineage empty. CLONE_UNTRACED was not audited; these results make no exhaustive lineage/history claim.

Before stop, independent checks found 341 recorded remote PID/start identities absent, four ports closed, both profiles absent, no runtime FD writers and the runtime lock free. All 123 remote outputs / 91,224,543 bytes were retrieved and independently reopened/hash-verified before delegated normal stop. Box bx_c26sux5t is ARCHIVED, current expiry NULL, initial expiry 2026-11-08T18:51:27.138Z. Normal stop stop_d1c77c185b83e0c3 completed 18:31:22.927Z, forced FALSE, error NULL. The independent observable 25-minute watchdog preceded the one persisted UUID allocation key; actual ID/expiry preceded runtime. Local lifecycle roles exited and their supervisor has no remaining children. Local drain and final independent artifact checks are in R.

Git SHA, relevant binary diff, small owned patches, provider pins and existing module/source path/hash references provide recovery. No source-tree archive, media duplication, peer WIP, old box resume/delete, deployment or subworkers were used. The selected generation is prepared for further causal testing, not qualified for gameplay input; the exact remaining cause is above.