JezzBall and Invisible War backlog audit — 15 October 2026

Superseding reconciliation — 18 October 2026

Checked against main e661285b2. The October 15 snapshot below is historical: its statements that IW transport and renderer candidates are absent from main must not be used to queue duplicate implementation or integration work.

Current source contains lib/process-send-broker.js with receiver ownership and retirement handling, and lib/d3d9-backend.js contains colorless-pass and independent RGB/dead-alpha handling. Both 3fdb29373 (combined native validation record/fixture repair) and 9eb485acb (renderer validation record) are main ancestors. Adapted implementations mean patch-unique historical commits alone are not evidence of missing behavior.

Combined native results record eight passing arms; renderer results record 77 observations per WebGL version and a discriminating negative control. Those existing receipts establish their scoped conformance, not current original IW playability. This reconciliation ran no tests or games and did not re-review raw screenshots. General pointer-message marshalling, cooperative transport, current original-game evidence, audio/device/FPS remain separate obligations. JezzBall's completed fixes remain completed; no repeat implementation is queued.

For GL/D3D work, Darkstone's bounded draw trace cffd439e2 is already merged and needs actual capture. B&W2's missing historical raw trace still prevents a causal renderer fix; bounded provenance instrumentation is now assigned to /root/bw2_layout_observer. The exact historical 27-app membership list is still missing; do not replace it with a guessed list or claim the goal complete.

Historical October 15 snapshot (superseded where noted above)

Source-only reconciliation against 5a31fcce8055e8385dba8ed76f6457767143068b. Remote main was subsequently verified at 5f43d41f54950151f08ad7cab3b536d9ed150bbd; the intervening six files concern CMR3 registration and color-key documentation, so none changes these findings. No build, guest, browser, device, or remote sandbox was run. Shared TODOS/STATUS and shared index were not edited.

Both JezzBall crash rows remain done. Invisible War remains held: the named lookup, process-lifetime, alias, transport, and rendering candidates are still outside main, with no patch-equivalent cherry-pick found. Historical gameplay on the held combined source is not current-main qualification.

The storage block is stale: this audit measured 2,588,237,824 available bytes locally, above the 2 GiB floor. This does not grant runtime concurrency; the sole runtime is owned by the separate kkrieger lane.

Exact unmerged source and checks

git merge-base --is-ancestor <commit> <pinned-main> returns 1 for each candidate below. git cherry <pinned-main> <candidate> marks each +, not patch-equivalent -. This is stronger than testing ancestry alone; it is not a guarantee that later unrelated code could not implement similar semantics. Direct current-source inspection confirms the relevant missing paths.

Main src/09a-handlers1-user.wat FindWindowA still calls the local find_window_core/window-table Z-order walk. Main host.js still registers _stopChildren to stop live children and uses child records in process_ctl; the candidate _processAliases/_duplicateProcessHandle path is absent. lib/process-send-broker.js is absent from main. Main lib/d3d9-backend.js does not have the candidates' colorless specialization or independentArgs/ sourceAlphaBlend dead-alpha elimination. No merged equivalent was identified.

The held broker registers one endpoint/stack per PID. Its canEnter rejects a request whose chain differs from the current owner, allowing only active descendants. That solves nested same-chain ownership but leaves independent cyclic waits held, and serializes unrelated receiver threads process-wide. The next focused regression should start independent sends in opposite directions, require progress without entering overlapping frames on one receiver thread, and separately show that two distinct receiver threads can progress. Design thread-owned admission/wait reentrancy before changing the gate; simply removing ownership would reintroduce overlapping callbacks. Retain retirement, deferred readiness, nested LRESULT/stack restoration and failure-path tests. No test in this paragraph was run during this audit.

The ledger's failure-collapse warning needs precision: e4510f1a9 already contains a held correction and documented two-Worker UD2/transport/disappeared- window regressions. It remains +/unmerged, so it does not clear the main integration hold. Do not repeat that fix. General pointer-message marshalling and cooperative transport remain unsupported; successful scalar WM_USER does not settle those contracts or the independent-cycle defect.

Retained evidence versus unavailable raw artifacts

Durable historical IW narrative is recoverable exactly with git show 09258c84368517e437e8b47b1a203789f25fc9ea:docs/re-notes/deus-ex-invisible-war-demo.md. Main's copy ends at the 10 October Reset investigation, so it must not be mistaken for the latest held-branch findings. The held narrative identifies 20261011T1157Z-iw-send-runtime-original, failure propagation 20261011T1216Z-process-send-failures, and scene/movement captures 20261011T1345Z-iw-rgb-game / 20261011T1357Z-iw-focused-input. The latter describes click, W, Escape tutorial dismissal, then W six seconds changing the corridor to the grating with HUD retained. HWND0 keyboard logging alone is not evidence of broken delivery.

Those named IW run directories are absent from this host's current scratch/runs, as are 20261011T1130Z-jezzball-worker-close-d10ba697 and 20261013T1225Z-jezzball-level-final. Earlier statements that they were copied locally describe the historical session, not this recovered workspace. This audit did not hash or visually re-review missing raw evidence. Recover exact bundles from an existing retained backup if a fresh evidence review is needed; do not recreate a result.json implying the originals are present.

Later local JezzBall folders are present: 20261013-jezz-next-level (108 files), 20261013-jezz-v86-dialog-compare (14), dialog geometry before/after (3/8), and 20261013-jezz-fps-measurement (6). Counts are inventory only, not reviewed gameplay, device parity, or hash verification. Main's RE note reports 301 GDI publications/15 seconds against the game's 50 ms timer, explicitly not a quiet-host benchmark; Safari/v86 cadence remains unmeasured.

Validation for this audit: commit ancestry, patch-equivalence classification, direct source inspection, local run-directory inventory, remote main receipt, and git diff --check. No rejected diagnostic was retried, delegated or rephrased. No gameplay or deployment conclusion was added.