Disciples II public input-pump boundary, 2026-11-17

The one authorized diagnostic original reached the genuine Search for Timmoria introduction through five confirmed setup presses, then stopped at Continue. The new public trace proves that its DOWN remained queued behind an unresolved native hover barrier until transport cancellation removed it. It does not establish why the target thread failed to acknowledge an exhausted pump. No map action, turn consequence, or native Quit was reached. The endpoint is permanently closed; no subsequent ticket query, capture, rebind, replay, or second original was attempted.

Evidence: scratch/runs/20261117-disciples2-public-input-pump-progress/result.json. The predecessor remains documented in the corrected hover report; its missing timing cannot be retroactively reconstructed.

Published observation route

Material commit d3f497035efb467a8389baf1fd0f66e54f20846a adds opt-in WINE_PUBLIC_INPUT_TRACE=/absolute/new-file.jsonl and an authored native fixture. lib/cli-public-input-trace.js observes existing host queue/ticket objects, renderer poll entry/return, native pump idle notifications, cancellation and independent release. It assigns object identities to queued events and records monotonic and wall timestamps. It reads no guest memory, invokes no additional guest API and changes no input scheduling, validation, deadline or receipt rule. Disabled tracing installs no poll wrappers.

The trace caps tickets at 64 and rows per ticket at 512, retaining a prefix and latest tail with overwrite counts. Queue snapshots include at most eight event headers. Observation ends at the immutable ticket deadline plus 1750 ms; file writes occur at lifecycle boundaries and exit. Hover generations are serialized as strings because native wnd_lookup_generation returns BigInt. The initial serialization failure and its failed qualification remain in the run directory.

The original used exact published descendant 3fcccea75aaaeb3f8201f8975684627a63014ba4, source manifest cca8ae93d0622f05dbe3d3ba6104bec61ec3f861a8e8d5fc4bcce948d0037f7d, module a105f6ce02932b3edebe9376a8b5eb4658f2159c52bb3cc76ae5ac4eadb05b0c. All 4585 Git bodies and inventory body hashes were independently verified. No foreign uncommitted overlay was adopted. Git and retained patches are the source archive; there is no source-tree snapshot or media bundle.

Native qualification and cost

The authored PE runs the production WAT and an actual CPUWorker helper. Its public window title distinguishes UP after DOWN from UP WITHOUT DOWN. Independent release is armed before DOWN. Normal, delayed-65-ms and delayed-250-ms cancellation cases passed against the sealed published module, with actual runtime/watchdog waits and both streams at EOF. The cancellation control retains the exact pending DOWN identity, removes it on UNKNOWN, and delivers the distinct UP identity without a late DOWN. This proves cancellation semantics, not the cause of the predecessor original's failure.

Seven trace-on/off guard pairs preserve consumption/refusal behavior for delivery, focus, owner, deadline, transport, handler exception and missing edge. Ring and BigInt cases passed. Five paired empty-poll samples measured added cost below 0.6 microseconds per poll. Published native maximum trace flush cost was below 1.36 ms. The diagnostic original recorded 18 flushes, 8.561 ms total and 1.387 ms maximum, with no sink error. These are observed costs, not universal bounds.

The mandatory canonical flock build and focused guard, transport, pointer mapping, native scene context and current-quest tests passed before material publication. Earlier slow-pump and cold-fetch failures remain preserved. A native 250-ms pump delay produced a 1137-ms DOWN confirmation; this demonstrates an authored delayed pump, not causation in either original. Exact unchanged proofs were reused rather than repeating broad qualification.

Actual original boundary

Media preflight checked all 128 existing files before native launch. The pipeline used shipped defaults, real ticks, no f.sg or imported save, a prearmed independent 3620-second watchdog, fixed 3600-second native lifetime and 300-second Quit reserve. Every intent followed personal review of a complete fresh image with a new nonce. The five setup confirmations were 57.0, 98.1, 61.2, 83.0 and 202.1 ms; their independent physical releases completed in 142–161 ms.

Continue at displayed (963,620) mapped to native (780,522), HWND 0x10001. The reviewed public window record identifies its thread as 1; the invisible CMqThread HWND 0x19001 belongs to thread 4. Relative to ticket preparation:

Time, ms Public observation
11.912 Host handler entered.
16.950 Handler returned with MOVE 21, hit-test 22, DOWN 23 queued.
33.337 Renderer returned MOVE 21; hover became native-pump/pending, not polled.
46.267 Renderer returned hit-test 22; DOWN 23 remained queued.
48.881–490.061 38 subsequent renderer polls returned empty with DOWN 23 still queued and hover polled:false.
505.216 Cancellation purge began with that exact DOWN 23 still pending.
505.241 Purge completed with no queued DOWN.
510.073 Transport disconnect recorded sticky UNKNOWN; cleanup UP 24 was queued.
528.473 Renderer returned distinct UP 24.
1311.489 Independently armed physical release completed.

There were 40 renderer poll returns before purge and 35 native idle notifications, all for thread 4. None acknowledged target thread 1. The retained ticket has 239 rows, no overwrites, and no DOWN-return candidate. The queue was actively polled; this is not evidence of an unpolled FIFO or a handler falsely counted as consumed. Purge follows the unresolved barrier and disconnect; it is not established as the original scheduling defect. Host poll rows do not identify the calling native thread or native Peek/GetMessage arguments/results. Continuous native messages, filtered target polls, thread routing and other reasons for the missing target acknowledgment remain distinguishable hypotheses, not findings.

Changing the barrier to accept a sibling's idle notification would discard the existing thread/window ownership condition. No such repair, synthetic CONSUMED, late DOWN, timeout widening or validation bypass was made. A follow-up can use retained logs and authored native fixtures to distinguish the target thread's public Peek/GetMessage filtering and exhaustion behavior. This worker has used its sole diagnostic original and cannot run another under this authorization.

Shutdown and retained images

Terminal UNKNOWN triggered file-based supervisor cancellation with no further interactive HTTP; the independently prearmed release completed as recorded above. Native, independent watchdog, supervisor child and holder all recorded actual exit 0 and both log streams at EOF. These are host shutdown receipts, not native Quit. The retained module was hash-verified before the holder's normal stop. All owned runtime PIDs and the listening port are checked offline after closure. No boxes were created: ID null, state NOT_CREATED, expiry null, stop not applicable.

Only six named intent-review images remain: main menu, New Quest menu, scenario, default race, default lord and genuine introduction. Seven loading/transition images were omitted with hashes recorded. There is no frame series, capture bundle, archive or gameplay screenshot claim.