Age of Wonders II: current browser movement and joined callback, 2026-11-05

Task: AOW2-RENDERING-COVERAGE-20261009. Evidence: scratch/runs/20261105-aow2-joined-callback/.

Current original observation

One original browser session reached Scenario → Single → Start, dismissed Julia's introduction, selected the six-unit army and completed two plotted and confirmed moves. I personally reviewed the current scene before each input and the before/selection/path/arrival captures. The army moved from the tower base to the southeast clearing and then south; the Party movement display changed 20/20 → 13/20 → 10/20. army-selected.png, plotted.png, arrived.png, second-plotted.png and second-arrived.png retain the visible positions, paths and changing panel. before.png retains the initial world. The first attempted selection hit a Gold resource; its selected.png filename does not prove army selection. army-selected.png is the actual selection.

Full captures and lossless, unchanged-pixel *-map-patch.png and *-party-panel.png extractions are retained. patch-extraction.json records the exact source hashes and crop rectangles. Earlier generic *-panel.png crops cover the right-side minimap, not the Party movement display. Two explicitly named reference-native-* images are recovered historical references, never new browser evidence. No historical coordinates were replayed without inspecting the current browser scene.

Mountains, trees, tower, army sprites, path markers and the Party panel render. Large ground areas remain black in the current browser images; full terrain parity is not established. The accepted native army/rendering result stands. No speculative terrain repair was attempted. The next rendering discriminator is the owning terrain/ground draw and its texture/blend/target state compared with the accepted native image, rather than another unchanged original run.

Completed Worker data attests actual legacy WebGL ownership at render endpoint 1, not merely a renderer request. Main slot 0 completed-fence sequences advance 3739 → 5032 → 5775 at selection/first arrival/second arrival; second arrival reports 473,339 draw calls, zero renderer errors/fallbacks and zero completed software/fallback draws. The reported context is WebGL 2.0 with renderer string ANGLE (Intel, Mesa Intel(R) UHD Graphics 620 (KBL GT2), OpenGL 4.6 (Core Profile) Mesa 23.2.1). This is the reported context identity, not physical-GPU qualification; launch arguments still requested ANGLE/SwiftShader. Full counters and endpoint options are retained in the checkpoint JSON.

Logical gameplay FPS remains unknown: there is no measured game-update cadence. Movements establish updates, not an FPS rate. Audio remains unqualified: no output was captured and headless Chrome included --mute-audio.

First original rejection and causal repair

The first joined rejection was retained synchronously from the production error-details log at 21:33:36.613Z, before retirement, even though its caller caught it. first-rejection.json identifies:

The original's main completed sequences continued through 9055 (Scenario), 10415 (Single), 20781 (loaded), 22199 (army selected), 22895 (first arrival) and 23401 (second arrival). This rejection was not a fatal stop. Miles thread exits/late slices are not established causes. The earlier synthetic releaseMainMessagePoint interleave is not substituted for this original issuer/endpoint/frame/epoch evidence.

The host's former root release read yield 17, cleared the incoming-send flag and cleared the yield through three separately awaited export requests. The original proves that the last continuation can reach an owning callback's scoped I/O instead of its suspended root message call.

The repair makes root message-point release one receiving-endpoint operation shared by actual Workers and the local endpoint. Any callback frame or outgoing root send returns an explicit callback-owned defer before accessing exports. Otherwise the endpoint checks the current yield is 17 and clears the incoming flag and yield synchronously. An ordinary different yield returns an explicit not-message-point result. Transport and retirement failures propagate instead of being swallowed. Existing export fences, wait reservations, scoped loader/ yield ownership, resume receipts/acknowledgements and retirement ordering remain unchanged. No guest state is inspected through a guard bypass.

The new real Worker-0/local regression fails against the old implementation (atomic-before.log): the stale root release reaches scoped I/O and silently returns no defer result. It passes after the repair and verifies scoped I/O, nested scopes, completed-but-unresumed callbacks, a callback itself at yield 17 without a reservation, root-send preservation, root restoration/ordinary release and transport/retirement error propagation. Existing diagnostics remain tested using an explicitly historical three-request transport probe; that probe is no longer described as the repaired production method.

original-order-negative.log additionally loads the original production host from the recorded Git commit in memory only and admits the scoped I/O callback specifically after the incoming flag clear, before clear_yield. That exact-order negative control fails; the same final regression passes against the repaired host (original-order-positive.log). No shared source was reverted, archived or replaced to run this control.

There was no second original session. Original movement/rendering evidence uses the pre-repair runtime; the causal repair is covered by protocol tests, not claimed as post-repair original-game acceptance.

Identity, validation and lifecycle

The ordered exact FINAL runtime AND file RELEASE ALLBUILDSFINISHED records were Aranna passive-resource2005 at 20:54:13Z, Conan first-draw2030 at 21:18:34Z, then legacy alpha16-target2040 at 21:29:42Z. Independent checks at 21:30:33.802749Z passed for 510 absent PIDs, 203 refused ports, 68 absent profiles, four free locks, zero heavy runtime and the disk floor. All shared edits/tests/build/browser/guest waited for this gate; the cleanup deadline is 22:15:33Z. Source-only preparation rehashed all 1,064 retained fixture pins in place without media copies.

Original EXE remains 5,748,736 bytes, SHA-256 a10590e5dbd013d154b00ea53e66670f4e74d38ab33adb2523f0a662af8f89f7, under test/binaries/win98-games-a-d/Age of Wonders2 demo-SW/extracted/. Current boot/fonts/native DLL closure, manifest and Worker-import prerequisites passed. Static imports are an inventory, not invocation or API compatibility.

Actual original source commit was 989ce3d323c40d40adbc3d86d134537d013962c5. The prelaunch source records agree; 162 served JavaScript responses match their recorded source hashes. The actual HTTP-200 noncached standard Wasm response is retained and hashes to 16dc14ec794aa225ac9cfd414b64049e4b7a8dc7e161585ac20584cc29ce00b6 (2,174,170 bytes). Built compat identity is 83b3bb3ddc4bf3efe8af8e99d6489a2cfafa0f3c210ba39633c2e0941499ed37 (2,176,816 bytes); the original browser used standard. JavaScript-only repair does not change these interpreter modules.

The sole browser ran 21:32:06Z–21:39:12Z using Chrome 151.0.7922.108, Node v24.18.1, a fresh owned profile, defaultViewport:null, and origin http://127.0.0.1:34955/index.html?debug&threads&d3d-renderer=webgl with COOP/COEP and cross-origin isolation. Actual generated production guest Worker URLs, full Chrome arguments, serving/CDP ports and completed slices are retained. The driver performs ordinary DOM mouse inputs and reads existing host metadata; it makes no guest export calls, debug setters, breakpoint changes or forced guest execution. No queued action timed out or was replayed.

Four focused protocol suites passed before the original and after the repair: reservation, scoped yields, frame ownership and wait retirement. The actual synthetic fixture is retained, 1,372 bytes, SHA-256 4e3e7a797017fda64615819b6f7160ef7a3aea0f197726f76e9dd6838f8036c2; it tests production JS protocol, not the full original interpreter. Final neighbor checks cover scoped loaders, wait identity and process-send admission/ runtime. The reviewed callback census retains general ownership as pending. Required locked full builds and exact command outcomes are retained.

An initial final-build attempt correctly failed a missed census hash update; its correction and later validation are retained. A corrected build's wrapper terminated with signal 143 while its detached compiler continued: the pending PIDs were resolved absent and complete module/data-check output retained before a separately supervised final build. The original driver's enclosing wrapper also returned 143 after ordinary closure; the Chrome exit code is 0, the driver wrote CLOSED ordinary stop, and its own exit status is unknown. No cause is invented for these wrapper terminations. No platform rejection/flag occurred.

Exact Git commits, relevant binary diffs, owned new-file patches, served-runtime identity, fixture pins and unique logs/captures/modules preserve recovery; Git is the source archive. No source tree archive, repeated media, download or extraction was made. No corpus candidate exists for this app, so evidence stays in this investigation instead of being attached to an unrelated dashboard card. No boxes, subworkers, public deployment or Heroes II session were used. Box ID/state/expiry: N/A; no box lifecycle operation occurred.

Final publication and independent process/port/profile/lock cleanup are recorded in the run's published.json and cleanup.json. All owned files/runtime are released only after those checks. Next: inspect the black-ground rendering discriminator; any new original run needs separate authorization and a material change. No logical FPS, audio or full terrain-parity acceptance is claimed.