NFS3 rehearsed attachment: local native timer abort blocks allocation
Run: scratch/runs/20261108-nfs3-rehearsed-consumption/.
Task incomplete. Canonical staging and actual production Worker attachment
now pass locally. The full local native route does not: the original displays
getcpuspeed - INITTIMER MUST BE INITIALIZED TO MEASURE CPU PERFORMANCE.
No box was allocated. Installed race bindings, active player/car consumption,
and input-caused race movement remain unmeasured. No emulator defect or fix
has been established.
Staging and attachment evidence
The released helpers from the preceding attempt were reused as run-owned inputs. Canonical relative-path validation rejects absolute paths, traversal, duplicate paths and byte mismatches. The observer was explicitly required from its intended run path in a separate Node process. Removing that actual file makes the production browser driver exit with ENOENT before it creates its runtime directory or starts Chrome or the original. This replaces the preceding failure mode of discovering the missing file after attaching a paused Worker. Worker attachment failures also close the runtime immediately.
The shared build had changed. The rehearsal instead served the retained pinned
modules, with published Git blobs supplying changed production JavaScript.
Both local2/module-7ec49….wasm and local3/module-7ec49….wasm, obtained through
CDP bytecode retrieval, independently hash to
7ec49cbbae4cb177909fd8fd9d3599713ee5f2a5ffb7f7e8c890848a540052bf.
local2/loaded-state.json associates that hash with each surviving original
guest Worker. The local3 bytecode file was collected during setup; the failed
game command did not produce its later Worker/module census. Its file alone
therefore does not independently prove original-Worker module equality for the
changed case. The runtime closure's published source commit is
0a38cf48438096482d059c5d8f27c048474c83b9; the starting checkout commit is
0c7a857e931d3a85c6bd4f081b78e3a2d4bd9f9b.
Generated x86 diagnostics on/off again matched all nine registers and 128 output words: the encoded Up binding returned 0 released and 255 held, and the original trace import ran exactly once per event. The exact eight-second Up+Left helper produced trusted Chrome events and independent reverse-order releases. A slow synthetic sampler preserved release timing and labelled its samples outside the hold. These bounded controls do not establish whole-game passivity or input causation.
The extended passive snapshot includes complete queue packets, availability words, sequences, simulation snapshots, playback selector, player pointer and car input-index/callback fields. Those additions remain preparations for race measurement; startup snapshots cannot qualify a player or binding.
Executed local native route and first new boundary
Three serial local browser cases ran; no heavy browsers overlapped.
localcompleted catalogue registration and native Direct3D setup. Its three-minute window refused game launch before dispatch because the command could not fit the existing cleanup reserve. Chrome was closed; the controller subsequently completed its immutable-cap cleanup. This was not transport UNKNOWN.local2completed catalogue registration, native renderer selection, warning acknowledgement and setup OK, then launched the registered original. Seven commands received exact ACKs, including normal quit. Three live guest Workers reported image base0x400000and an installed diagnostic snapshot. A bounded diagnostic arm independently expired in all three Workers. However,loaded.pngvisibly shows the native INITTIMER abort. The console records it at 06:51:26.564 UTC. Thus successful attachment and release did not imply successful game startup.local3tested a changed diagnostic attachment order: Runtime enable, observer installation and Worker release occur before debugger enable; debugger collection is deferred until capture.attachment.jsonlandattachment-release-analysis.jsonrecord an exact release for every attached Worker. The original still reports the same abort at 06:54:58.251 UTC. Command number 4, IDgame, then exceeds its 20-second work deadline at 06:55:01.529 UTC. The driver closes Chrome with sticky UNKNOWN. No command was resent and no later runtime was launched. The added internal eight-second observation wait was also inside that work deadline; the variant is a failed experiment, not a qualified remedy.
local2/loaded.png, both consoles, local3/first-failure.json,
local3/transport-closed.json, attachment receipts and process-drain evidence
preserve the new boundary. No missing-EXE, endpoint or focus explanation is
asserted.
Source inspection identifies the abort's zero test of guest 0x562ae8 at
0x4f044b. The initialization routine can clear that field on several failure
paths, including a timer-readiness timeout. This is not a measured execution
trace of those branches. Prior initialization/API return values, the specific
clearing path, and a matched observer/debugger-off original control were not
captured. In particular, the displayed text does not prove attachment caused
the abort or that a zero clock reading was mistaken for an initialization
sentinel. No clock, configuration, scene or guest state was forced.
The next required work is a separately bounded local timer-initialization experiment recording the actual predicate, prior API returns and readiness counter against a matched observer/debugger-off control with the same loaded module. The full native local route must pass before a fresh watched original box measures binding → packet → queue → simulation → selected car controls.
Preservation and shutdown
Evidence remains local in the owned run. base-commit.txt,
relevant-source.patch, runtime pins and individual new-helper patches retain
recovery information; no source/media archive or media copy was made. The
final helper patches describe the final failed candidate; per-case startup
manifests were not separately retained across local helper revisions, so they
must not be presented as complete byte-level recovery of every earlier case.
The retrieved browser modules and unique screenshots/logs are retained and
hash-verified.
All three original-route Chrome processes exited 0. Controller/Chrome PIDs, ports, profiles and writer locks were checked absent; the generated control browser also exited 0 and removed its profile. The messageboard watcher was stopped. Disk remained above 5 GB. No box/watchdog was armed because allocation never became eligible: box ID none, state not allocated, expiry N/A, stop receipt N/A. No inherited box was resumed, stopped or deleted.
Only this note is committed. No production emulator code changed. Full27, parity, audio and unmeasured FPS remain open.