Shadow Man corrected capture technical review
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.
REQUEST CHANGES for capture usability. Exact corrected input
aac3cf5db81fc2946f9bfc2f09a11a9c7be021ec resolves the eager-arm,
compile/Module-binding and scoped census/discovery defects in review
43a1d1d7b8c0418cdc1d61f646e2ea5150a318c4, but its recipe still requires
fresh-run values before they can be supplied through the current interface. It is not ready for a separately authorized full-module/guest capture. No production repair is included in this acceptance branch.
Demonstrated blocking gap
The preload parses SHADOW_NULL_SELECTION before compilation and passes that
object at instantiation to createSelectedSend. That function copies the four
integer keys into a private frozen object. There is no setter, selection
registration control command, deferred selection provider, or supported way to
read environment changes back into this object. Instantiation occurs before
the runner installs its interactive control channel.
The actual runner increments diagnosticActionSequence as each scheduled input
is dequeued; it uses the current execution batch, then resolves modal HWND or
renderer/slot HWND and dialog parents. It invokes the selected-send hook before
the synchronous we.send_message. Its ordinary input receipt is appended only
after the send returns. The corrected ordering is good, but it exposes the
fresh resolved values only to a hook already committed to pre-start constants.
Our executable combines the exact runner queue/dequeue prefix and exact
dlg-cmd branch, the actual preload and compile/Module binding, and the
entire actual sender body with the previous review's scripted dependencies. Small imported window resolvers supply the synthetic route; this is not an
invented replacement action driver that calls the hook with a known context. The actual branch constructs that context and calls the native sender itself.
fresh-selection.cjs compiles the retained selection-path.wat fixture and
spawns each control with an actual 5-second external process timeout. Results
and exact child argv/environment/stderr are in fresh-selection-results.json;
per-control .trace and .jsonl files are retained.
| Control through actual branch | Actual exit | Evidence |
|---|---|---|
| Pre-known synthetic modal / renderer-parent / slot | 86 each | Pre-arm receipt, sender-pre, ordinary first-zero, terminal first-zero; no returned final-send log |
| Fresh modal / renderer-parent / slot HWND | 1 each | wrong or late selected action/HWND; no arm or event record; source count remains zero |
| Actual execution batch 20 for action scheduled at 19 | 1 | Same rejection before send |
| Set environment selection to fresh HWND after OK, before final dequeue | 1 | Trace records exact correct late values; private frozen selector remains unchanged |
For concrete reproduction, all cases start with synthetic
{actionSequence:3,batch:19,cmdId:3,hwnd:8194}. Configuration1024 and OK1
return at batches17/18 with diagnostics still OFF. Fresh controls resolve
HWND12291 for the selected action instead. The env-update control supplies
{actionSequence:3,batch:19,cmdId:3,hwnd:12291} before dequeue, yet still rejects. These numbers are fixture controls, never suggested guest constants.
The desired fresh-selection assertion also actually exits1 (check9):
node -e "const r=require(process.argv[1]);for(const c of r.results)require('node:assert/strict').equal(c.exit,86,'fresh selected send must acquire and accept current HWND/batch via actual runner interface: '+c.mode);" /home/user/wine-assembly-shadow-man-null-arming-independent-accept-oct23/docs/re-notes/shadow-man-null-arming-independent-accept-20261023/fresh-selection-results.json
The ordinary harness exits0 because it confirms the rejection; the desired acceptance oracle exits1 because current source does not satisfy it. The three normal controls succeed on the identical Module and extracted source path.
Fixed schedules can determine an action sequence and intended batch in advance. An overdue queued action still runs at the actual current batch. The fresh HWND is the decisive missing input, and interactive late batch selection has the same immutable-selector problem. Existing freeze/snapshot/eval controls can inspect a live run; that does not reconfigure this private selector. Changing the global hook or adding a custom arbitrary-JS adapter would be new capture code requiring review, not an executable recipe supplied by this candidate. No guest run was used to infer future values.
The retained two earlier receipts establish only returned sends, not successful OK, a particular launcher screen, persistence or incident ownership. Separate observable route evidence remains required. The report's instruction to choose fresh values after verified OK currently conflicts with its immutable pre-start environment interface. Restarting would create another run; there is no supported pause/rearm selection update in this helper.
Corrected conditions accepted within offline scope
GIT_NO_LAZY_FETCH=1 python3 docs/re-notes/shadow-man-null-arming-independent-accept-20261023/run-checks.py
- Both actual modal and renderer/parent/slot branches call the hook after fresh resolution and before the send. The no-hook path sends directly; absent dialog sends nothing. Deliberately moving arm after send fails the ordering assertion. One-shot context/action/PID/runner/command/HWND/batch checks reject before selected execution; duplicate/late arm also rejects.
- Actual sender Configuration/OK plus40 actual modal-completion prefixes remain OFF and consume zero source/host event records. Selected arm resets counters. Source event33 traps; a returning observer cannot resume a first-zero event. The host reserves one of32 records for terminal zero (the block record counts; pre-arm receipt does not). Quota starvation, returned selected send, exception, retention failure and required-block-before-zero fail terminally.
- Compile copies actual bytes, compares them before native compilation, rejects start sections, then binds the returned immutable Module privately. A changed valid custom section, substituted bytes, unbound Module, byte-form instantiate, constructors, streaming and repeated compile/instance contexts reject. Start sections reject before execution. The normal Module control captures. The old6d145e9 helper still falsely succeeds at86 for eager startup and changed Module; both updated-contract assertions actually exit1. These tests use dedicated fixture drivers, not full-runner/guest identity qualification.
- All records copy registers/stack/code synchronously, retain sealed immutable
values and fsync before process exit86. Missing exports also exit86 but carry
capture-failed; exit86 alone never means success. Wrong selector rejects with exit1 and a missing-terminal failure, rather than inventing capture.
The ordinary CACA0005 continuation intentionally writes EIP=0. The actual sender
fixture reaches first-zero before its restoration and labels it
ordinary-or-unknown; zero-writer unproven. We accept that classification. Neither zero nor the decoded block proves the historical incident or a zero
writer. A required-block policy must reject zero-before-block.
The actual full sender body covers OFF caller restoration and64-round abandonment: sender-pre follows saved registers/child allocation, event2 precedes identity/depth checks and restoration. Saved EIP/ESP are caller values; current token identifies the child, not proven modal ownership. The exact modal sentinel/completion prefix verifies event3 after teardown/reentry and sentinel adjustment but before consumed-frame reads; event4 after previous pump and EAX/EIP/ESP restoration, before ended/result/quit reset. Saved ESP names the consumed frame, not the original API entry. Scripted teardown mutates private pump fields to distinguish those positions. This does not execute the whole dispatcher, interpreter, guest or canonical module.
Census, discovery and unchanged source
The discoverable .js wrapper is unit tier, with75-second registration and
three actual20-second child suite guards. Scoped timeout validation passes;
full test manifest remains pending. The exact pinned fragment checker passes
all three WAT files, which remain byte-identical to6d145e9. Source-only export
declarations total1054 without duplicates; this is not an actual Module export
inventory. Inputs, hashes and exact four changes are in source-checks.json.
Native fixture and qualification limits
All fragment memories have fixed initial=maximum128pages, 8MiB, no full guest heap. Compiler executes in place from the review donor with recorded identity; it is the focused test compiler, not qualification of missing full candidate vendored provenance. Modules:
The new fixture adds only resolver imports/exports and a sender alias to the retained3394-byte source. Resolver, translator, invocation and callback/run dependencies are scripted. The native sender and CACA0005/zero predicate are actual source. VM extraction executes the actual runner branch/dequeue prefix, not runner bootstrap or a full guest route. Each new child has a5s external watchdog and VM calls1s; suite commands have75s external guards. Wasm arena limits do not bound total Node/compiler RSS. No full handler/module/runtime, canonical export compatibility or performance qualification is claimed.
Build and guest prerequisites remain separate
The exact existing pinned recipe was read and identity-checked, not executed:
/home/user/wine-assembly-shadow-man-null-hook-independent-review-oct23/docs/re-notes/shadow-man-null-hook-independent-review-20261023/pinned-build.sh. Review43a1d1d and repairaac3cf5 list its exact gate order. After the selection
gap is repaired/reviewed, a separately admitted build still needs complete
src/main.watx include closure, full compiler provenance, mirrors/baselines,
full callback/transitive census, test manifest/timeout registration, all
fragment/region/memory/API/dispatch/layout/import/provenance gates, and
include-ordered paren/label checks. No cheap gate is waived by scoped30 PASS.
Only then separately authorize vendored-WATX full compilation, data overlaps,
distinct revised size/SHA and actual Module imports/exports/start absence. Common required exports: nine GPR/EIP getters, guest_read8, set_bp,
set_trace_eip_range. Revised mode also requires set_null_diag,
null_diag_field, null_diag_token, null_diag_root. Qualify real sender/modal
paths and pinned runner/import topology against that module. No canonical
module was obtained or inspected here. Historical pre-block identity remains
1,879,387bytes/SHAccd5b0a8e4fcb84d06bf539c8a4af28b377472072f3422750122048e9339f553,
separate from every revised diagnostic fixture/module.
Guest authorization must separately supply the complete retained game fixture (partial479-file VFS is invalid), target process/PID and runner identities, fresh observable OK/launcher evidence, usable one-send selection, explicit required/optional block policy, exclusive output and finite external watchdog that can kill a synchronous Wasm spin. Historical batch/HWND constants cannot fill this gap. No guest command, runtime/media recovery or lifecycle operation was attempted. Persistence/reopen/Play/gameplay remain untested.