Disciples II native hover consumer boundary, 2026-11-17

Authored native fixtures distinguish a filtered owner pump from an exhausted unfiltered pump. They establish proper refusal under the existing hover contract; they do not identify the cause of the previous original's missing thread-1 idle. The prior original and its endpoint remain closed. Evidence is in scratch/runs/20261117-disciples2-native-hover-pump-completion/result.json.

Public native fixture findings

The authored PE first establishes an actual owner-thread unfiltered idle, then switches through an ordinary WM_APP message to a window-filtered or range-filtered PeekMessage loop. An actual CPUWorker independently pumps its empty queue. MOVE and NCHITTEST pass through real USER and DispatchMessage. Once the native hover barrier is armed, filtered owner queries and sibling empty queries cannot finish it. DOWN remains queued; 500-ms transport cancellation purges that exact edge, and independent UP reaches the real wndproc as UP WITHOUT DOWN. An unfiltered control reaches UP after DOWN. PM_NOREMOVE followed by removal preserves the same ordering. An earlier fixture that switched before readiness was established did consume DOWN; its failure is retained, not treated as an armed-barrier proof.

Production-WAT public-thunk tests exercise both filter forms and both removal modes, actual x86 MOVE/NCHITTEST callbacks, filtered-out message retention, filtered GetMessage's native wait, sibling refusal, and the legitimate transition back to owner unfiltered exhaustion. The complete public API transcripts match with native observation enabled and disabled. Wrong thread, memory, generation, replacement window, retirement and hidden target cannot acknowledge hover. Actual nested x86 PeekMessage inside a wndproc retains a separate observation identity. Existing guard tests cover focus/actor/context changes, cancellation, release-before-delivery and late DOWN refusal. No timeout, actor, target, frame, focus, writer, receipt or independent release rule changes here.

Essential diagnostic

The old public trace cannot identify the native consumer or its Peek/GetMessage filters at a host empty poll. set_user_pump_observation(budget) now enables at most 4096 observations in that WASM consumer; the default is zero. Peek/Get entry records include native thread, MSG pointer, HWND/range/removal filters, stack and return address. A separate handler-return identity records EAX, yield reason, stack position and a returned MSG header when its API frame was consumed. Disposition 1 means consumed API frame, 2 means yielded/parked, and 3 means another stack shape, including possible callback redirection. Dispatch entry retains the actual public MSG HWND/message and its own sequence. These counters are diagnostic identities only; routing and readiness never consult them.

WINE_PUBLIC_INPUT_TRACE samples those exports at the existing renderer poll and ticket lifecycle boundaries. It identifies the bound actor/main source and records the target's actual native owner/generation, memory match, renderer-window identity and retirement/visibility. It invokes no additional guest API, reads no private game state, and does not run another guest slice. Unsupported or retired diagnostic exports produce an unavailable record while preserving dequeue and refusal behavior. Consumer entry and last completed handler have separate IDs; their fields must not be combined as one call when the IDs differ. The explicit callbackReturn: unmeasured is intentional: neither host dequeue nor native handler redirection proves that an application callback returned or processed a frame. This is sampled public execution, not a complete API-call history.

The existing limits remain 64 tickets, 512 rows/ticket and deadline plus 1750 ms. An idle-tail qualification exposed loss of the distinct UP in the old rotating ring. At most one quarter of those same rows now preserves exact input and lifecycle identities, alongside the prefix and rotating tail. Lifecycle rows retain their latest native sample so cancellation's filter/wait identity survives later idle traffic. Overflow counts remain explicit. No queue action or native acknowledgement is inferred from those records.

Native on/off public-API samples and actual CPUWorker controls measure the cost; raw sample/flush timings and earlier failed tests remain in the run. Observation budgets exhaust without changing public return values or callback execution. Seven trace-on/off guard negative pairs and unavailable-export failures preserve the original outcomes. Canonical and exact published-module qualification are recorded with the final outcome in the evidence result.

Original and next boundary

The one fresh diagnostic original used published 7810e8a2a0f8cef8bb27daa956453a62c2934382, Git-only full source SHA 714208d3b7ec4e7f2f89ce6619c593a8b617330ea0d8d8cb3c847d18c8b607ea and qualified module SHA a01678daa31393ac24eff152c7ac9586700eeb166ef7dcbe20e642faa2434a2c. All 4589 Git bodies were independently verified; no owned patch or foreign WIP was applied. The intervening published commit changed another worker's report only; the sealed runtime inventory remained identical. Main, actual Worker and child inheritance passed. The canonical flock build and focused guard/context tests passed. Exact sealed native units, trace units, ordinary unfiltered control, filtered on/off controls and PM_NOREMOVE control all passed with actual wait/EOF receipts. The ordinary wndproc saw UP after DOWN; both filtered controls saw UP WITHOUT DOWN, with no late DOWN. The 128-media preflight passed unchanged.

Five individually reviewed finite setup clicks were consumed and independently released. The personally viewed complete fresh quest-intro frame is retained as original/questintro.png; it is a story introduction, not a map. Continue then reached permanent UNKNOWN at the unchanged 500-ms transport boundary. There was no subsequent interactive HTTP, rebind, replay or second original. Only the already armed independent release continued.

The new public observations materially narrow the missing fact. During this specific Continue attempt, the same thread 1 owns target HWND 65537, generation 1, with matching window and memory and a live visible target. Its consecutive USER entry identities are 16939 through 16943: prior WM_TIMER dispatch; unfiltered PeekMessage(..., 0, 0, 0, PM_REMOVE) at +244.276 ms; NCHITTEST dispatch at +244.936 ms; another unfiltered PM_REMOVE Peek at +466.494 ms; and SETCURSOR dispatch at +468.840 ms. Times are relative to trace preparation. Completed Peek identities 16940 and 16942 each returned a message (NCHITTEST and SETCURSOR), not exhaustion. The last observed entry at cancellation remains SETCURSOR dispatch. These positive observations refute a filtered-query or changed-owner explanation for those calls. They do not measure application callback return, prove that an application frame processed, or explain why owner exhaustion did not occur. Later owner Peek entries and timer/message dispatches show continuing public execution after cancellation; they cannot authorize delivery of the purged DOWN.

Exact DOWN 23 remained behind pending hover and was purged at +499.718 ms. Distinct UP 24 returned at +503.218 ms. Independent release acknowledged at +1500.239 ms. Retained pre-purge rows include at least 127 renderer returns, 125 null returns and 121 sibling idle entries, all observed on thread 4. The 512-row per-ticket cap was not increased: final ticket saw 700 rows; the union of bounded flushes retains 683, with sequences 129–145 absent. Counts are lower bounds; positive USER entry identities above are consecutive and retained. No absence-based claim relies on those missing rows.

Original native observation cost was 1976 host samples, 26.954 ms total, 0.293 ms maximum; 18 bounded flushes cost 13.118 ms total, 3.115 ms maximum. Published filtered native control measured 1026 samples, 27.976 ms total, 0.942 ms maximum and 4.657 ms total flushing (2.208 ms maximum). Raw contention outliers from earlier qualification remain in evidence; these are measurements, not universal scheduling bounds. On/off public transcripts and seven guard negative pairs remained equal; observation adds no guest calls or slices.

Native, independent watchdog, both outer wait owners and the module holder all returned zero with actual waits and both EOFs. The unique module was retrieved and hash-verified before normal holder shutdown. The control port is closed; own processes and watcher are absent. No boxes were created. No map action, turn consequence or native Quit was achieved. No contract defect was established and no speculative readiness repair was made. The next boundary is actual native callback/consumer completion and legitimate hover readiness/wakeup under this observed unfiltered, nonempty owner path, using public execution evidence. This original stays closed; another original needs new authorization.