Additional actual-browser startup crash sweep, 2026-11-08

Evidence: scratch/runs/20261108-graphics-additional-browser-crash-sweep/ (ignored). Scope is Tomb Raider II demo, Tomb Raider III demo and Drakan demo, each under cooperative/threaded scheduling and Software/WebGL selectors: twelve intended cells across three titles. This extends the named six-title startup sweep; it does not establish historical 27-title membership, gameplay, backend parity, FPS, audio quality or full playability.

All twelve complete-fixture cells reached native startup dialogs without a classified crash or deadline. Their owning graphics backends were not created.

Runtime and method

The coherent published-main source is 78299c7e97460eab8ea422e169dfcab75b6a525b. Tracked runtime inputs were read as individual Git blobs at that commit; runtime-pins.json records each source commit, Git blob, byte size and SHA-256. Peer-owned NFS3 worktree files were neither edited nor reset. The relevant worktree diff was empty when pins were composed (excluded-peer-WIP.patch); using published Git blobs for every tracked runtime path also excludes any later peer edits. Relevant untracked runtime inputs and owned helpers are retained as patches. Git remains the source archive; there is no source-tree or media archive.

The standard module is 2,182,500 bytes, SHA-256 f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. registered-apps.json retains the exact published registrations. The authentic registered closures contain three TR2 paths, ten TR3 paths and 64 Drakan paths, including its manifest and all registered resources. Shared DLLs and standard system data are separately pinned by existing path, size and SHA-256. Guest executables, media, configuration, working directories and registered startupRegistry are unchanged.

The released native CDP driver, actual Worker/module recovery, static server, passive owner observers and composited PNG capture are reused. Both ordinary renderer selectors are set together, with threads or no-threads. There is no gameplay/dialog input, extra guest/config/clock/scheduler override or CLI substitute. Each case retains the 60-second setup bound, four five-second startup observations and immutable 100-second browser cap. Capture and later diagnostics have separate timestamps. An independent observable detached 25-minute watchdog was armed before the UUID-idempotent single fresh box allocation; its deadline was never extended.

Local validation before allocation checked every enumerated pin, syntax, observer delegation and zero input, two sequential actual Chromium cleanups, and the supervised normal-stop/watchdog lifecycle with a local mock. Existing classification, actual browser owner/shadow/later-thread boot and DirectDraw4/7/Surface3 descriptor regressions passed. No production code was changed, so no new implementation build is claimed.

Preserved staging correction

The inherited staging list omitted the existing optional test/binaries/tlbs/stdole2.tlb. Initial TR2 threaded Software, threaded WebGL and cooperative Software executions reached their bounded startup checkpoint, but log its missing system-data HTTP404. These three original executions remain as setup-gap observations and do not count as complete-fixture acceptance. The serial parent was terminated without signaling its active browser, which then drained normally. The authentic 15,088-byte local typelib was hash-pinned and staged. No guest bytes or runtime implementation changed.

Corrected passive executions use distinct -complete names and retain all twelve intended cells. Original and corrected pin lists, transition records, raw logs and captures are preserved. There was no input replay or first transport UNKNOWN replay. The initial preallocation check covered its enumerated 667 pins; it did not detect the inherited list omission. The corrected list has 668 pins.

An added serial descendant assertion initially split process command lines on a literal backslash-zero. Its chromeChildrenAbsent field is therefore not accepted as independent proof. Parent PID, ports, profile, writer and command checks remain separate; an independent process watcher uses actual NUL splits, and the terminal full-command-line audit supplies final descendant evidence. preallocation-correction.json records this limitation without changing old receipts or claiming a runtime repair.

Results

All twelve corrected intended cells completed. Across them, no classified escaped guest/Worker trap, unimplemented API, unexpected guest/browser/target exit, setup failure, nonreturning call, transport UNKNOWN or case deadline was observed. The three incomplete initial executions remain separate, explicitly classified setup gaps. There are 15 executions, three titles and twelve complete intended cells; none of the intended cells is missing or unexecuted. No demonstrated emulator regression warranted an implementation repair.

Title Personally reviewed boundary in all four corrected cells
Tomb Raider II Native Graphics property sheet, Primary Display Driver, Hardware 3D Acceleration selected, Z Buffer, Full Screen 640×480 High Color, OK/Cancel. No Venice scene.
Tomb Raider III Native playable-demo welcome, Lara portrait, Important Notice and OK/Cancel. No later setup, title or Jungle scene.
Drakan Native Beta Demo Warning, beta code 95% complete and OK button. No logos, menu or world.

Three ordered four-cell sheets and the three original setup-gap captures were personally reviewed; representative original-resolution PNGs confirm dialog text. All 15 original PNGs pass CRC/IDAT decoding. capture-reviews.json binds each review to its exact PNG hash. Raw console, startup/slice observations and exception/observer records were checked separately from pixels. Drakan's four ole32 built-in fallback warnings are retained; they are not observed unimplemented-API traps. Generic cleanup text pending: UNKNOWN; never replay is template wording, not an actual transport outcome.

final-review.json records every corrected cell independently. Each has four startup observations, zero input, a stable capture owner and the expected actual CPU placement: a guest Worker for threaded cases and local boot for cooperative cases. Capture requests occur 20.002–20.008 seconds after startup; capture/diagnostic overhead remains separately timestamped. Applicable page and Worker COM registry identities agree, with 82 entries and distinct Surface4/Surface2 entries. Actual instantiated standard modules match the f965 digest in all 15 executions. All 2,327 complete HTTP reads and 53 loaded host-source reads match their staged pins, with zero identity mismatches.

Every captured case has no type-20 D3D device, no render endpoint/owning render Worker and no live GL context. Owning graphics API, context and backend are therefore explicitly not-created. CPU software flags and page bridge backend options are retained raw, but neither requested selectors, the checked guest Hardware 3D option nor GDI presentation establishes an owning graphics backend. This passive sweep supplies no active Software/WebGL rendering or parity proof.

target-lifecycle.jsonl, actual process inventories before/after observation and Chrome exit records retain the supported process/target evidence. All Chrome exits are code0/signal null after intentional close, and no target crash or owning-target destruction before close was observed. These records do not invent an exact page-to-OS-renderer-PID mapping or native crash reason. The independent watcher collected 4,317 process samples, covering the entire corrected serial, with no simultaneous original profiles and no remaining processes. The final audit checks all 668 pins unchanged, all 15 controller/ Chrome pairs absent, all HTTP/CDP ports closed, profiles/writers/commands clear, both serial parents and the process watcher absent, and no owned live runtime.

Retrieval and shutdown

All 521 remote artifact references, 424 unique payload hashes, were retrieved and independently reread for byte size/SHA-256 at 2026-11-08T05:27:59.858Z, before normal stop. Equal payloads can share local hardlinks; each recorded artifact path remains independently verifiable. Unique logs, original captures and actual modules are retained. Recovery uses the Git SHA, source/helper patches and existing-media references in RECOVERY.md, without source-tree or media archives.

Fresh box bx_uvsbnjzm is archived, stopped at 2026-11-08T05:28:00.546Z. Final expiry archiveAfter is null; initial expiry was 05:44:12.505Z, with the unchanged independent watchdog deadline 05:44:12.345Z. Normal stop receipt stop_185e4cf474b04b2e completed at 05:28:00.572Z, forced=false, error=null. Normal and watchdog delegates exit0, and the supervisor records no remaining children. Sampled local free disk remained above five billion bytes; continuous minimum disk was not measured. No RAM gate, old403/security-task retry, old-box resume/deletion, public deployment or nested worker occurred.

Only this note is committed. No production code or guest binary changed. Next work requires separately scoped ordinary native dismissal of each setup/ welcome warning before observing an actual owning backend or gameplay. Prior TR2/Drakan gameplay findings remain separate evidence; this startup sweep adds no new gameplay, historical27, parity, FPS, audio or full-playability claim.