Aranna atomic observation, 2026-11-07

The incomplete-publication failure is reproduced with captured failed-read bytes and corrected in the run harness. The transport helper and focused regression are published on main as 344c54a2cebb0af51858c80b78c18e665dc52080. Fresh composed pointer qualification stopped at a generated-control targeting blocker, so no fresh original was launched and world movement remains unverified. Evidence is in scratch/runs/20261107-aranna-atomic-observation/.

Demonstrated transport cause and its limits

reproduce-publication.cjs executes the retained queue reader and a separate Node publisher using the old writeFileSync(finalPath, Buffer, {flag:'wx'}) protocol. The test pauses the publisher immediately after exclusive open, before its first payload write. The reader sees the final pathname, reads zero bytes, and throws SyntaxError: Unexpected end of JSON input before validation, admission, adapter preparation or receipt. The actual returned bytes are captured at the reader call, with length, base64 and SHA-256 in reproduction/failed-read.json and publication-reproduction.json. The empty read SHA is e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855. After releasing the publisher, the final command is complete valid JSON. The failed iterator is never retried.

This proves that making the final name visible before completing its contents can produce the observed exception. It does not recover the failed-read bytes of the predecessor's original command9. That historical cause remains an inference. Command9 remains terminal UNKNOWN, with no reinterpretation, replay, resumption, or zero-effect claim.

tools/atomic-command-publication.js writes and fsyncs an exclusively created temporary sibling, closes it, then publishes an exclusive hard link to the final name and removes the temporary name. Incomplete temporary contents are never visible at the polled command name. Competing writers cannot overwrite a final record: losers receive EEXIST. This is a same-filesystem immutable-publication protocol; no power-loss durability of the containing directory is claimed.

The new run-local queue writer uses this helper for reviews and commands. The reader uses it for ready/admitted/done records, and its command read captures the exact failed bytes if parsing fails. A malformed final record still throws immediately, without admission, retry, repair, or replay. The supervisor's deadline Quit publication also uses the exclusive helper. The retained queue validator, production input adapter, owner/module/review/geometry/protocol/point guards, quotas, UNKNOWN closure and stale-refusal conditions are unchanged. No emulator or game source behavior was changed.

Verification

node --test tools/atomic-command-publication.test.js passed all four tests: paused old publication reproduces the empty read; paused atomic publication keeps the final name absent; eight simultaneous publishers yield exactly one complete immutable observation; malformed final records preserve exact bytes and refuse replacement. The 256 KiB concurrent records exercise substantial payloads rather than only tiny synchronous writes.

transport-regression-v2.json records six checks against the actual retained legacy reader and the exact transformed new reader loaded by this run's driver: empty and prefix records fail before validation/admission/preparation; complete wait/screenshot observation and Quit produce valid receipts; UNKNOWN prevents the next admission. The earlier six-case regression is also retained, with its narrower test scope. Mapping controls passed six valid/five invalid cases; selector controls passed nine zero-effect refusals; 32 route-contract boundaries preserved UNKNOWN/partial closure and the stale zero-effect recovery rule.

flock /tmp/wine-build.lock bash tools/build.sh passed before the explicit-path code commit. The commit was immediately pulled/pushed with strict SSH settings. The generated baseline modules still match the shipped standard and compatibility modules. Actual Worker passivity controls passed all twelve baseline/disabled/armed cells across both modules and ordinary/Worker routes. Both-module ordered broker controls passed signed seek, short read, EOF/error, immutable identity and actual production Worker checks.

Preparation mistakes remain recorded: the first passivity invocation reached the identity check before compatibility observer preparation finished and failed before its browser execution; the later completed-module invocation passed. Initial broker preparation used an unsupported output CLI argument and did not emit the requested module; its browser then rejected HTTP "not found" bytes before guest execution. Preparation with the documented absolute WINE_RENDER_ARTIFACT_PREPARE path and the subsequent broker control passed. These failed preparations are not original attempts.

First composed qualification blocker

Lords2 explicitly released runtime/source/build at 03:08:41Z. The independent handoff check at 03:09:45Z verified five PIDs absent, two ports refused, one profile absent, no matching descendants and all four locks free before taking the lane and claiming the two new implementation files.

The standard generated full-shell control acquired actual screen lock and armed the unchanged owning observer. The actual owner was slot0, endpoint1, target 61A6F0149F021B27F554F7728997A390, session 8B9149B76CE65BCC11256B994548BFAE, observer SHA e7d6617cecdf3275db05454763c07cad7e6eb557afc2a14ca3413d960328fd31. The requested relative movement at command2 was (2,-145) from the measured native (397,300) toward the reviewed Single Player target. Owning manager stores first reached (399,155) at serials968/971, then applied the same displacement again to (401,10) at serials980/983. Its completed receipt is timestamped 03:14:22.442Z. A subsequent generated diagnostic correction (-1,72) accumulated once, reaching (400,82) at serials988/991, with completed receipt at 03:15:25.692Z. There was no target activation. The source of the variable repeated delivery is not established here; no emulator defect is inferred. qualification-blocker.json retains the stores, owner, receipts, native readouts, image dimensions and hashes.

This fresh control did not qualify the composed route. Compatibility full-shell qualification, expiry/stale refusal, stable unlock, neutral click, typing and reacquisition were consequently not completed in this run. The predecessor's successful controls and proven explanation of the old (392,599) mismatch remain valid evidence for that predecessor. Its guest accumulation and repeated DOWN/UP displacement correction was retained without weakening any guard.

All five generated commands, including normal Quit, have complete receipts; there are zero UNKNOWN/partial commands in this control. No original EULA, Choose Hero, name entry, Normal, map, world, or movement was attempted here. The coordinator was notified on the messageboard at 03:15:55Z. No further guest-input operation followed the diagnostic correction except normal Quit.

Visual review and identities

Screenshots were inspected at their original 1001x800 resolution. Startup and response02 show the generated Single Player rectangle. Responses00,01 and03 have black guest GDI areas; no disappeared-control inference is made. Responses 00 and01 are byte-identical, SHA 9b8707e2974692c3d2f4e278038bfc04481825c89d6c64a2e7ecc1fe4028202f. An inaccurate response00 review was written alongside the image tool call before its pixels were assessed. The original note is retained, with explicit response-00.review-correction.json; its claim that the rectangle remained visible is withdrawn. Later observations do not inherit that visual claim.

The standard baseline is 2,181,179 bytes, SHA 07f1cc970038f8ab31313b8b200edf26caa63473f293f5fc937f5cfe5c5c7f6e. Compatibility baseline/observer SHAs are respectively 1d4bd4330ea16947959e3f4472339d5a32a61216cd8b5246198d11046b652c3e and 6b5edb202b5e03d84056b8663869f8485de2b5a7d592ece0e350dc43a88a005d. Actual Worker instantiated modules, HTTP identities and passivity results are retained. Runtime source pin v2 and post-runtime checks record unchanged production/runtime inputs, with only the offline regression subsequently expanded. Final pin v3 includes those additions. Exact Git commits, tracked binary diffs and relevant untracked/new-file patches provide source recovery; there is no source-tree or media archive.

All 28 existing game media paths passed length/SHA-256 checks against the predecessor's retained identities. Guest media/configuration/save bytes were not edited. Media was referenced at existing paths without copying.

Cleanup and next step

Normal generated Quit completed at 03:15:55.657Z; driver and Chrome exited zero. The lock holder exited normally after its release marker. At 03:18:15Z, final-cleanup.json independently verified all 29 recorded PIDs absent, six ports refusing, three profiles absent and four locks free, with 3,066,933,248 available bytes. These counts cover retained lifecycle records; the early failed broker preparation's lifecycle filename was reused by its later control, so these are not claimed as an independent census of that earlier browser. Its failing supervisor exited one and its finally block closed the browser. Runtime/source/build and implementation-file ownership were explicitly released at 03:18:22Z. No boxes were assigned or created: ID/state/expiry and box shutdown handoff are not applicable. No automated cybersecurity risk flag was observed.

The next task is to establish the generated relative-delivery discrepancy and complete both-module composed pointer qualification before any separately authorized fresh original. World movement remains unverified. Findings are published with this note; result.json is published last, then this worker stops. Own pending-result schema, artifact containment and existence checks passed. The repository-wide evidence checker still exits one for the unrelated retained Alice run's outside-directory reference; its report is preserved without changing that agent's files.