Shadow Man fresh selected-send repair — 2026-10-23

Historical source last updated, 2026-10-23. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.

Supported selection contract

The existing preload now accepts this explicit action intent through its existing SHADOW_NULL_SELECTION variable:

{"mode":"selected-action","actionSequence":3,"cmdId":3}

3 is the synthetic fixture's dequeued action sequence, not a guest-run recipe. Choose the one intended action sequence from the actual input schedule, counting all dequeued actions. The intent allows only these three keys, requires a positive safe-integer sequence and final Beenden command 3. HWND and batch must be absent. The helper privately copies/freezes the intent. It still validates PID, runner path and byte pin, exact instantiated Module/instance, dlg-cmd, sequence, command and supported resolved-dialog route. This selects one explicit action; it does not accept every send or a later action with a matching command.

At the existing actual runner pre-send boundary, after modal/parent/slot resolution, the adapter validates current nonzero u32 HWND and safe-integer nonnegative current batch, and snapshots the four resolved fields. The existing capture arm path deep-copies/freezes, seals, writes and fsyncs that exact context with the original intent, returned prior-action receipts, pre-arm registers, source count and metadata before breakpoint/trace/diagnostic mutation and native send. One-shot use is consumed before arming; retention/arm failures cannot rearm. Returned sends and exceptions still fail terminally.

Executable acceptance

GIT_NO_LAZY_FETCH=1 python3 docs/re-notes/shadow-man-fresh-selection-repair-20261023/run-checks.py

Exit 0, ten guarded checks. Exact argv, cwd, environment deltas, exits, stdout/stderr and 75-second external command limits are in checks.json. The discoverable existing unit entry now executes four focused children with 15-second external guards, within its unchanged registered 75-second timeout:

GIT_NO_LAZY_FETCH=1 \
SHADOW_NULL_TEST_COMPILER=/home/user/wine-assembly-shadow-man-null-hook-independent-review-oct23/tools/watx.js \
SHADOW_NULL_TEST_OUTPUT="$PWD/docs/re-notes/shadow-man-fresh-selection-repair-20261023" \
node test/test-shadow-man-null-capture.js

Exit 0. The full sender and modal-prefix fixture is reused unchanged; the new test extracts the actual runner queue/dequeue prefix and actual dlg-cmd branch, not an invented action driver. runner-extract.js records those bytes. The new driver draws a fresh HWND in [12000,60000) and a current batch in [20,30) only after both synthetic Configuration1024/OK1 sends return. No future HWND or batch appears in the preload intent. Resolver values are scripted, not real guest observations. Native sender body/CACA0005/zero predicate are actual source.

fresh-selection-results.json and per-case .trace/.jsonl retain:

Case Child exit Assertion
Fresh modal, renderer parent, slot 86 each Actual differing batch and fresh HWND acquired after OK; exact receipt before native sender
Fresh modal plus environment replacement 86 Original explicit intent remains privately selected
Wrong command, stale exact selection, stale exact + late env update 1 each No arm/event record and no final send receipt
Absent dialog 0 Runner reports NO DIALOG; no arm, source count0; missing-terminal record correctly prevents capture claim
No hook/default OFF 0 Direct returned native sends; source count0, no capture
Invalid mixed intent/late HWND key 1 Rejects at adapter creation, before dequeue

Every positive receipt seal is verified; pre-arm source count is0; first native sender record count is1. The fsync observer records the actual sealed receipt only after the real fsync returns, before the first native sender record. fresh-unit-controls.json independently exercises actual capture with instrumented exports: persist → breakpoint/trace → arm → send, deeply frozen intent, mutation-resistant private selection, duplicate/rearm rejects. Twelve wrong PID/runner/instance/action/sequence/cmd/invalid HWND/batch/resolution contexts and nine invalid registrations reject without arm/send.

The original maintained suites also pass: actual sender Config/OK +40 modal completion prefixes consume zero quota while OFF; source event33 traps; ordinary zero cannot continue; terminal reservation, exceptions, retention, returned-send, required-block and starvation controls remain accepted. Compile-byte copies, changed valid Module/custom section, unbound Module, constructor/streaming/alternate forms, repeated compile/instance, start section, wrong PID/runner hash and missing exports retain their expected rejection. Exit86 alone remains insufficient: a missing-export capture failure also exits86.

Old candidate rejecting controls

Before editing, the review's exact actual-runner/native-sender reproducer ran against the old candidate: preknown modal/parent/slot exit86; fresh modal, parent, slot, differing current batch and late env update exit1 before send. reproduce-old.cjs retains that reproducer and reexecutes against the read-only exact old candidate in check6, exit0 because rejection is its expected finding. old-candidate/fresh-selection-results.json retains exact input IDs, argv and outputs. Check7's desired fresh acceptance assertion actually exits1 on those observations. It is not reported as an acceptance pass.

Check5 executes the new desired fresh fixture against the exact old runner and preload using SHADOW_NULL_TEST_SOURCE_ROOT and isolated old-intent-oracle output: exit1 at expected86 versus actual1, with invalid final Beenden selector. The same new fixture accepts the repaired adapter. Source-root override exists only in the test fixture; production helper has no such bypass.

Pins, census and limits

New selected-send native fixture: 3526bytes, SHA256 103b6e04fc7b5a30a820eba0d76c171fa8f4db5967860849ad9fb89c4a44a00e. Retained sender/modal fixture:3394bytes SHA256 5b874002a970b97d246408739811bd1802a8c31a6dc2357973579ce67c55f8ab; original hook:902bytes SHA256 f45414f005d33707d05e0a16d9abf6d77f2ddce5aa00ddcddc73551e68e500dd. Each fixed initial=maximum128-page arena is 8MiB. Resolver/translator, invocation allocation and callback/run seams are scripted. VM calls have1s limits; each fresh/binding child has5s external timeout. Arena limits do not bound Node/compiler RSS. WAT and recovery commands are retained; ignored wasm outputs are reproducible and their hashes recorded. No full/default/canonical compile, guest, media, Boat, browser, GPU or runtime/recovery was attempted.

Next prerequisite and handoff

This closes the offline fresh-selection execution prerequisite only. Before native module or guest qualification, independently review the adapter and complete/admit the existing exact source closure/provenance/manifest/transitive census/gates, then separately authorized canonical compilation and actual imports/exports/start absence inventory. The retained pinned build recipe in review43a1d1d remains unexecuted here. Qualify revised module with actual runner, PID and single instance, complete game fixture, observable fresh OK/launcher route, the explicitly counted selected action intent, exclusive evidence output, block policy and external watchdog. No historical HWND/batch can substitute. Configuration/OK returned receipts do not prove visual success, persistence, reopen, Play, gameplay, modal ownership or a zero writer. Ordinary CACA0005 first-zero remains ordinary-or-unknown; zero-writer unproven.