Stronghold first native coverage invalidation

The released original d2d4cfa9940d3dde7a62fd3ecc3a32eb8c7913d1 reached Choose Type / Custom Game hover, then failed before Custom Game DOWN with profile 2, image 0x400000, epoch 346889932, callback ABI 1, reason 5, depth 1. Its first invalidating site/time was not retained. That historical failure and its seal are unchanged. Setup, world and hostile engagement were not established by that run.

Evidence for this continuation is scratch/runs/20261105-stronghold-first-native-invalidation/. Canonical media is referenced by existing path and SHA-256, without copies: test/binaries/candidates/stronghold-crusader-demo/installed/App_Executables/Stronghold Crusader multiplayer demo.exe, 9024c533095f2227b0105b7e8a9e4cf28ec4fb206b165629787b2c97cf99f13f. Git SHA, relevant binary diff and owned new-file patches provide source recovery; there is no source archive or staged source tree.

Preparation remained scratch-only until keyedmip1230 FINAL at 13:40:29Z and Lords1310 FINAL at 13:48:35Z, both ALLBUILDSFINISHED. The independent release gate checked 17 recorded processes absent, four profiles absent, seven refused ports, four free locks, no heavy runtime and over 9.6 GB free disk. D3D's four early browser exit statuses remain unknown; resource absence does not supply normal-exit evidence. Lords1310 removed its unvalidated implementation and released without any original launch after an ordinary repository census review gate. No successful Lords build is inferred. Shared ownership began 13:56:55Z.

The implementation adds only an immutable first native coverage failure record. Existing reason stores pass through a recorder that copies state before the store. The synchronous marker import runs after the store, so an import error cannot leave coverage valid. The owning observer installs its marker hook before instantiation and captures UTC and the actual Wasm caller/offset, independently of whether block observation has started. Later failures, setters, callbacks, thread/epoch checks and arm attempts cannot replace the record. No coverage predicate, target set, callback ABI, guest control flow, readiness or pixel acceptance is relaxed. No causal emulator repair is claimed.

pending_context_first_failure_version() is 1. pending_context_first_failure(field) returns an unsigned 32-bit word when read by JS; out-of-range fields return zero. Reason zero means no retained failure. The record is diagnostic and never constitutes coverage proof. Its fields are:

Fields Meaning
0–3 first reason, predicate site, guest EIP, guest ESP
4–10 current thread, armed thread, depth, observation epoch, profile, actual image base, active-run bit
11–15 callback entries/returns, compiled PC/EIP, yield reason
16–19 rejecting predicate's local a/b/c, reason before the store
20–27 exact eight words of the top native frame, or zeros without a frame
28–31 current native process epoch low/high, invocation root low/high
32–35 dispatcher local return, thunk index, metadata, current thunk EIP

Site 0 uses the retained native caller/offset. Sites 1–6 identify, in source order, the rejecting predicates in pending_entry; a/b are its PC/ESP. Sites 101–108 identify those in pending_dispatch_callback; a/b/c are entry EIP, return EIP and parent ESP. Dispatcher locals remain distinct from the redirected callback EIP and the saved top frame. The diagnostic schema is also retained in the run's diagnostic-schema.json.

The focused generated regression passes all 14 cases across the two shipped modules: untracked target entry, tracked-call stack escape, host setter, bad RET, foreign epoch, diagnostic import exception during execution, and an unarmed control. It checks immutable native/host records with zero observed blocks, pre-store state, actual native stack, fail-closed consumer rejection, later events and getter bounds. The first test attempt incorrectly entered a callback after switching to a foreign thread; that existing ownership trap was retained in the failed log and the test order was corrected. No emulator workaround was introduced for it. Accepted startup prerequisite qualification was not repeated.

The exact census scanner reports no changed inventoried producers. Its file is unchanged; the lifetime review records scalar copies and synchronous forwarding without guest context, token, lease, scheduling or continuation changes. Existing callback ownership remains pending. The named current runtime binding is dated for this assignment and the two exact predecessors; historical bindings remain closed. Command/review/native/input/pixel limits remain unchanged. The full flock /tmp/wine-build.lock bash tools/build.sh passed. The actual composed-entry check passed with 2429 filesystem and HTTP rows, all 45 negative controls, zero Chrome/original launches, and server closure. The single finite original observation completed below.

No box is assigned (ID/state/expiry N/A); coordinator owns box lifecycle. The task delegates normal teardown of its own local runtime. No subworkers or public deployment are used.

The diagnostic was committed and published as c11e415e02f0dfbe6f7f38a18a5d0fa5e2833ef7. Explicit strict-SSH pull/push passed; an initial configured pull used HTTPS, which was corrected by using the explicit SSH URL. The first launch command at 14:04:07Z failed closed on a stale Git identity before runtime preparation, release issuance, browser creation or original-attempt creation: concurrent published 2b2e9ac48 changed only TODOS.md. Both manifests and that refusal remain in evidence. Binding the unchanged diagnostic to published HEAD 2b2e9ac48b764ed36833efcbdd8ea12199929f14 allowed the one original at 14:04:42Z. Its finite caller completed normally at 14:09:48Z after the first concrete command failure. No second original ran.

The first native invalidation is now established at 2026-11-05T14:05:22.292Z, during startup, before the later title/menu observation. The immutable raw first-failure.json has SHA-256 5e98b15702c02a40c469dc2d7b5a4e6c00e4d9bc4ed817bfe845c3d28b0c6a33. decoded-first-cause.json, first-failure-retention.json and the original attachment reports retain its interpretation and subsequent equality checks. The synchronous native stack is retained verbatim, including function indices 496/497/501/13884/14973 and offsets 0x1d78d/0x1d7b3/0x1da6f/0x1ded22/0x1e4821. The line-regex function-name utility does not account for this compiled WATX module's function layout; its attempted mappings are not used as evidence.

Captured fact Value
First reason / site / previous reason 5 / pending_entry site 4 / 0
Rejecting predicate unsigned current ESP greater than top-frame entry ESP
Guest EIP / ESP 0x0057c707 / 0x074fff20
Current / armed thread; depth 1 / 1; 1
Profile / image / callback ABI 2 / 0x00400000 / 1
Observation epoch 2574209027 (0x996f5003)
Native process epoch / invocation root 1 / 2
Top-frame target / return / entry ESP / kind 0x07500210 / 0x0057c707 / 0x074fff08 / 1
Top-frame parent EIP / ESP / token 0x0057c707 / 0x074fff20 / 0
Callback entries / returns 3 / 2
Current thunk EIP / active / yield 0x07500210 / 1 / 0
Compiled PC / EIP 0 / 0

Unmodified original disassembly resolves the call at 0x57c701 through IAT 0x584130 to KERNEL32!ReadFile; 0x57c707 is its immediate successor. The original pushes five arguments, with a NULL OVERLAPPED argument. The captured ESP delta is 24 bytes: return address plus five arguments. This is consistent with the documented x86 stdcall callee cleanup. Microsoft's ReadFile contract distinguishes synchronous completion from overlapped I/O; an internal lazy-fill retry is not itself a guest callback or completed read.

Source comparison identifies a concrete candidate: handle_ReadFile calls io_block for internal lazy-fill result 997. io_block restores the exact stdcall frame, redirects to current_thunk_eip, sets yield 12 and stops the slice. pending_dispatch_callback can record a redirected thunk as a kind-1 callback; ordinary API completion does not execute the guest RET path that changes that kind to 4. The captured kind-1 target equals the retained current thunk, and the current EIP/ESP equal its stored parent EIP/ESP. This supports an internal-retry classification mismatch, but the record does not include the preceding dispatcher entry, ReadFile host result, park transition or frame creation. Those missing facts prevent proving that exact lifecycle in this original, or certifying a coverage-preserving repair. lib/dll-loader.js's native-token DllMain path was inspected; this record does not establish a DllMain sentinel/token mismatch. No speculative callback exemption, shadow-frame pop, epoch reset or coverage relaxation was applied. The causal repair remains unestablished, so this task stops with the exact diagnostic retained.

All 21 screenshots were personally inspected, including all eight requested reviews before their deadlines. Ordinary controls reached title at 8448, main menu at 8576, Choose Type at 8768 and Custom Game hover at 8832. Command 10 failed before CustomGameDOWN on continuous coverage rejection. The retained first record remained identical in three owning attachment reports and in the terminal native getter copy. Native blocks/publications remain zero; ordinary setup, world and visible hostile engagement remain unestablished. No readiness, presentation or pixel certification is inferred from the screenshots.

closure-after.json verifies all 2429 sealed dependency rows still match, including the canonical original media and both built modules. Module SHA-256: 006b2e5adeb4d05e4647fd5581f87c4caeb3b2bc038fb4a945307be8240b2afe and compatibility c652f4eb35f380ffdb5f77ea1e210397028993e64ba726716f592862f7266753. No source archive, media copy or staged tree was produced.

Normal own-runtime teardown requested Browser.close at 14:09:48.047Z and retained actual Chrome exit 0, signal null at 14:09:48.295Z. final-cleanup.json at 14:14:08Z independently checks ten recorded processes absent, the own profile removed after no users remained, three refused ports, four free locks, no heavy runtime and over 9.5 GB free disk. The own watcher was already absent when cleanup began; its exit status is unknown and no kill was sent. No broad kills occurred. Box ID/state/expiry remain N/A. All builds and local runtime work are finished. The next causal work needs a generated ReadFile park/retry lifecycle reproduction and exact ownership proof before any repair or further original; this task performs neither a workaround nor another original observation.