Alice held-phase and relocation prerequisites, 2026-11-03

STOP: the first native instruction-fault qualification failed. The task-local adapter holds ordinary and cooperative-thread breakpoint stops, and the original DLL passes relocation-aware instruction checks after a real native loader run. However, a native fault still exits the CLI instead of retaining the earliest inspectable process/thread stop. Alice was never launched in this task. usableForFooting=false; no coherent authentic ground sample or movement claim exists. The known unsupported/airborne comparison remains needed.

Exact input source was 8365c62ee8f0f516210a2aafd3253425db5dd334. Only this note is committed. All candidate adapters, native instruction fixtures, commands, logs, original PE/assets, hashes and recovery files are retained in ignored scratch/runs/20261103-alice-held-phase-relocation-repair/. No production paths changed; no foreign WIP entered the runtime source closure. There was one worker and one fresh boat, without subworkers or successor work.

First failing native prerequisite

The failure occurred at 2026-11-03T16:49:59.992Z, after the preceding breakpoint and relocation cases passed. A disposable native instruction fixture loaded the original winmine PE, allocated its own instruction/stack spans, placed 0f 0b (UD2) at actual guest address 01022a74, and armed a later inc ebx instruction at 01022a77. It submitted one four-credit request.

Actual CLI PID 14844, session bc3b3b4e-9ded-4dd5-9f11-ad0dcdff9620, logged:

*** CRASH at batch 0: unreachable
EIP before batch: 0x01022a74
EIP=0x01022a74 ... ESP=0x01024a78
01022a74  0f 0b  db 0x0f, 0xb

The original RuntimeError: unreachable and WASM stack are preserved in native-fault-cli.log; they were not suppressed or converted into a successful boundary. The pending step lost its socket. A subsequent inspection could only report ECONNREFUSED, recorded separately from the original native fault in FIRST-native-fault-prerequisite-failure.json. native-fault-event.json retains the actual instruction address, process/session and first transport consequence. The focused fault fixture exited1. There was no retry, later fixture, original opening, route input or extension of a guest deadline after this failure.

The immediate repair boundary is concrete: the adapter calls its hold observer only after a successful run() return. An exception bypasses that observer and reaches the CLI's existing crash handler, which calls process.exit(1). A complete observer must latch the original fault and owning instance before that teardown, park every scheduler admission and preserve pending-credit state, while retaining the original error as failure. That path is not qualified here.

Why the retained authentic boundary did not remain held

The prior authentic capture was reproduced read-only with its retained reproduce-prerequisite.py and receipt29; independent PID/session and archive hashes passed. Its two failures remain separate: EIP 00453411/halt1 instead of 00452f63/halt5, and DLL 02d40000 instead of original 10000000. Sticky caller 02dd7aa2 corresponds arithmetically to 10097aa2, but establishes no held phase.

exact-control-source.json preserves the relevant pinned source extracts and command order. Commands27–29 were arm-by-eval at batch4104, step1, then eval at batch4105. Frozen credits account outer batches, not individual native run() calls. Within a credited batch the ordinary main run is followed by cooperative worker slices, with as many as three interleaved main runs under the default four-slice schedule. The native breakpoint sets bp_skip_once on halt; another run() at that EIP dispatches the stopped block once. No persistent scheduler latch connected that native stop to frozen control.

The post-run WASM breakpoint path around test/run.js:10716 logs/rearms; it does not await a scheduler hold. In this authentic session the breakpoint was armed through exports.set_bp() by eval, so that action did not itself populate the CLI's breakAddrs array. The original trace preserves subsequent movement boundary/return instructions, including a boundary with a fresh run budget. Only at the next control boundary does finishPreviousControlBatch() drain the one credit and resolve the step; receipt29 therefore observes the completed outer batch rather than the first native breakpoint return.

This explains the missing scheduler hold. The retained trace has no admission site labels, so it cannot identify whether the first hit was ordinary or an interleaved main run. No finer historical attribution is claimed.

Concrete task-local candidate and passing native cases

cli-adapter.cjs compiles an exact, hash-checked copy of the pinned CLI in memory; it does not edit production files. Its task-local ThreadManager transformation returns immediately when a cooperative native breakpoint is captured, before another sibling can run. held-gate.cjs captures actual exports, registers, process/session, thread census, batch and pending credits synchronously, then parks the async caller before later scheduler work. Read-only span captures use that same instance and epoch. Arbitrary eval, step, input and unfreeze commands are refused while held; inspection eval exposes only the saved held anchor. Explicit token release disarms once and grants no new credit. Capture errors retain the earliest breakpoint epoch and error and refuse release.

These are actual x86 instruction fixtures executed by the compiled WASM, not receipt labels or authentic Alice footing:

Fixture Measured result
Main, PID11005 inc eax; jmp reaches actual 01022a77, halt5; EAX1/EBX0, batch0 and all four pending credits remain held.
Repeated inspection Three snapshot/eval/read passes preserve registers, thread census and credits; native instruction bytes read 40eb0043eb00ebfe. Mutating eval and additional steps are refused.
Main release The original request completes exactly four credits; EBX becomes1. Duplicate token release is inert and leaves the exhausted-credit snapshot unchanged.
Capture error A separate breakpoint epoch rejects an invalid span, keeps its original native stop/error and refuses later reads and release.
Cooperative, PID13508 Actual Win32 thread ID2/cache slot1/handlee1000 stops at 01022a87, halt5, before the next sibling runs; both live thread EIPs and four credits remain stable through inspection. Release completes the original four-credit request.

The initial main adapter preceded the thread-target dispatch extension. The unchanged gate and actual final thread/fault adapter are retained; the initial main adapter version was reconstructed by reversing the recorded extension. adapter-versions.json states that provenance explicitly: there is no independent per-execution seal of the initial adapter. The production source and module are independently sealed. Worker-backed and nested callback scheduling are not qualified by these cooperative tests.

Strict original-PE relocation candidate

relocated-module.cjs requires the full original file hashes:

It reads the actual process DLL table and parallel lifecycle rows, including recorded path, live state, generation, image bounds, reservation/capacity and allocation token. It rejects ambiguous loaded spans and requires a unique fgame mapping. It compares complete mapped PE headers with the pinned original, including sections and relocation directory. Original signature bytes must match their manifests before translating RVAs; expected mapped bytes apply the PE's individual HIGHLOW relocation entries. Partial relocation operands and unsupported relocation types are rejected. No preferred-base fallback or blanket byte-delta normalization is used. Later translations recheck process/session and the actual mapping row/generation.

The native fixture instantiated the freshly built module, loaded the original engine, and used the existing loadDlls() path for the original DLL, including its normal DllMain. Actual DLL base was 01039000, lifecycle state2, generation1; DllMain returned EAX1 and initialization completed. 53 original signature windows and370 relocated operands passed. Original 10097000: b8 00 d2 1c 10 was actually mapped at 010d0000 as b8 00 62 20 01, returning mapped export-table address 01206200. This is real loaded-PE relocation evidence, not a synthetic module table.

Eight checks rejected changed DLL hash, wrong process/session, wrong module RVA, nonexecutable section, invalid cross-section/image span, partial mapped read and stale process epoch. Actual retirement/reuse of a lifecycle generation was not exercised; callback provenance, Player allocation lifetime and post-boundary sample rejection remain prerequisites for complete capture integration. The accepted selector24/decoder17 proofs and original manifests were reused unchanged and hash-checked. Their synthetic cases were not relabeled as runtime ground. The prior capture-expression.js and strict decoder were retained but not invoked against Alice, and a complete relocated live-ground decoder was not qualified before this task stopped.

Build, preflight and recovery

Before allocation, the FINAL polling bundle SHA256 b7aefa122818d5c353e3df59ef03b21081620b342f149530f8abd0cb1ae29410 passed verification, followed by16 protocol checks, four actual wrapper checks, and the delayed1.5-second single-submission zero-wait case. No former faulty runtime wrapper was extracted. All19 original/support/fixture pins, including notepad and winmine PEs, passed existence, size and full-SHA checks before allocation and again before build.

The committed source closure was sealed and streamed without foreign WIP. The boat-side locked full build and keyboard-message-lparam, frozen-stdin-control and CLI DI-mouse suites passed without skips. The compiled module was2089591 bytes, SHA256 0a58fd7b9ac82c5303048776f442b15e72528b55ddf85564e0df92fcb1aa384d; the relocation fixture directly instantiated those bytes. Production source and original fixture pins matched again during recovery.

Preparation limits are retained. Initial transport identity/template errors preceded native admission. After build and focused suites passed, the task-host import checker found an omitted launch-plan.json; adding that host file made the exact import preflight pass without a rebuild or guest-fixture retry. Native instruction fixture logs warn that winmine's optional original system DLLs were absent; those fixture programs executed no application APIs. They establish instruction/control behavior, not full winmine compatibility. The full build also retains its explicit absent-Git-HEAD API comparison and offline Big5 check.

All 5935 remote payloads were retrieved and locally checked for size/full SHA256 before normal STOP. Archive116514442 bytes, SHA256 b92f8cf42a6150963b7e717119264c91e484e7d9572fe7cf8f2da39a1d567beb. RECOVERY.md, archive-verification.json, payloads.json and SHA256SUMS preserve recovery. No PNG was produced by these instruction-only cases, and there is no new authentic game image.

Own boat bx_fj5qxw5h, original expiry 2026-11-03T18:08:59.729Z; ID/expiry and original-task delegated normal STOP were recorded before use. Normal stop stop_0d2bc64850586acd completed 16:53:23.697Z, error:null, forced:false. Final info confirms STOPPED, completed available snapshot verified16:53:23.637Z. Own runtime helpers were absent and port8123 closed before stop. Custody returns to the coordinator. No predecessor/foreign box or process, Claude pane, other game lane, heldCACA003B or public deployment was touched. Fixture-only setup writes were confined to disposable native test instances; Alice PE/assets/config/save bytes were unchanged, with no Alice guest execution, injected engine call, cheat, map load or movement.

This single task ends at the reproducible native fault-retention failure. The partial adapters remain uncommitted evidence; only this tested investigation note is published. Native fault retention must pass before a separately authorized original opening and coherent supported-ground capture. The supported sample, allocation continuity and explicit unsupported comparison remain needed. AUTHOR RELEASE; no successor or runtime reuse.