Shadow Man diagnostic arming and module identity repair
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.
Corrected behavior
test/run.js assigns a monotonic one-based action sequence when each scheduled
input is dequeued. Both actual dlg-cmd sends consult the explicit
Symbol.for('wine.shadow-null.selected-send.v1') interface. The modal path
resolves its native modal HWND first; the other path completes renderer/slot
resolution and parent climbing first. The hook then validates the actual
instance, PID, runner identity, action, action sequence, actual batch, command3,
unsigned fresh resolved HWND and resolution path before invoking the actual
we.send_message(dlg,0x111,ev.cmdId,0). No Wasm export property is reassigned. With no preload/hook, the original send executes directly.
Instantiation does not arm. Startup, Configuration1024 and OK1 remain OFF and
cannot consume host or source quota. The selected final Beenden send is a
single synchronous window. Earlier sends pass through while OFF; at most two
returned dlg-cmd receipts are retained in host memory, then copied into the
arm receipt. Returned means the synchronous send returned, not proof that a
setting persisted or a particular screen appeared. Before diagnostic mutation,
arm synchronously retains a sealed/deep-frozen receipt containing the fresh
selection, prior returned actions, GPR/EIP state and diagnostic fields. It then
clears breakpoint, selects only decoded address0x004015d4, resets host counters
and calls set_null_diag(1) to reset/enable source counters in revised mode.
The selected-send callback must complete by synchronous terminal capture. Return without capture, exception, quota/starvation, required-block failure or
retention failure is terminal failure. Rearm and duplicate/late selection are
rejected; wrong context/action/PID/runner/HWND/batch/command cannot send the
selected message. Captures remain immutable copies, synchronously written and
fsynced, followed by OS exit86. Exit86 alone is not success. A
capture-failed terminal or missing terminal receipt fails acceptance.
At most32 host event records are permitted, with one reserved for terminal
first-zero; the arm receipt is separate and does not consume event quota. Source event33 still traps. Early callbacks while unarmed do not record or
consume host quota. The normal source default remains OFF. A first-zero record
is explicitly ordinary-or-unknown; zero-writer unproven. A legitimate
CACA0005 zero can end the window before the incident. A required-block policy
retains that observation but terminates with
capture-failed: first-zero before required block. This is never incident or
zero-writer evidence. Earlier Configuration saved locals are not recoverable
from this final-send window.
The preload hashes the named file, copies and compares the actual compile input bytes before native compilation, and binds the returned immutable Module in a private WeakSet. Instantiation accepts only that Module and one context. Altered custom sections, other bytes, unbound Modules, byte-form instantiate, Instance constructors, streaming forms and repeated compile or instantiate contexts reject before the rejected context executes. Module section boundaries are parsed to reject start sections before instantiation; any import observer callback during instantiate is also a terminal failure. The actual Module therefore derives from the bytes in its identity metadata. This assumes the dedicated pinned runner and preload are trusted; it is not a sandbox against hostile JS retaining original constructors/functions before preload. No start-function compatibility is claimed or allowed.
Complete helper input contract (no guest command executed)
A future authorized launch must supply the real corrected test/run.js as the
entry script, one --wasm=ABSOLUTE_FILE argument, and these environment fields:
SHADOW_NULL_PID: exact target OS PID, established by the external launcher before exec; inherited child PID or worker-thread contexts reject.SHADOW_NULL_RUNNER_SHA: SHA256 of the exact corrected runner,4d7b9587676db38fcf3314807acb2b2a42b5f09d2c98fcf3dccb8a6cde9e1809. Its resolved entry path must be this checkout'stest/run.js.SHADOW_NULL_OUTPUT: absent exclusive output file in a retained destination.SHADOW_NULL_SELECTION: JSON with integeractionSequence(one-based, positive),batch(actual execution batch),cmdId:3, andhwnd(fresh, nonzero unsigned32). Choose these from fresh observable route evidence and the newly resolved launcher, after verified Configuration/OK. Test values sequence3/batch19/HWND8194 are synthetic controls. Historical HWND0x1001c and batch2578996 are not fresh-run constants. All scheduled actions count toward actionSequence; only actualdlg-cmdboundaries can arm.SHADOW_NULL_MODE: exactlyhistorical-pre-blockorrevised-first-zero. ObsoleteSHADOW_NULL_UNTIL_ZEROis rejected even if set to0.SHADOW_NULL_BLOCK_POLICY:requiredoroptional, explicitly chosen.SHADOW_NULL_WATCHDOG_MS: integer1..300000, matching a finite external process watchdog. This is a declared launcher contract, not proof a watchdog exists; event-loop timers cannot stop synchronous Wasm. The external launcher must retain timeout/no-capture failure separately. No watchdog/guest launched here; native fixture children had actual5-second process timeouts.- Revised mode additionally requires
SHADOW_NULL_REVISED_SHA: exact named-file SHA of a separately built and qualified diagnostic module, distinct from the historical module. Historical mode enforces exact1,879,387 bytes and SHAccd5b0a8e4fcb84d06bf539c8a4af28b377472072f3422750122048e9339f553, and uses no revised exports. No canonical module was available/executed here.
No competing trace/break/watch flags, --spawned-child, other selected-send
hook, alternate compilation/instantiation, propagated installer preload or
second context is permitted. The runner's ordinary wall/batch guards remain
useful but do not replace the external watchdog. There is no pause/rearm or
multi-phase capture contract. Coordinator must select a complete retained game
fixture/process; partial479-file VFS remains invalid. Prior successful OK and
fresh launcher selection require independent observable route evidence; two
returned send receipts do not establish those user-visible conditions.
Durable tests and reviewed census
test/test-shadow-man-null-capture.js is discovered by the exact .js regex
and conventionally assigned unit tier. It runs the original capture fixture,
the maintained actual-source sender/modal-prefix fixture, and repair controls. tools/test-timeouts.json registers75 seconds; each of three child suites has
an actual20-second process watchdog. Each preload control has5 seconds. No global discovery regex/tier convention was changed.
Only four census entries were corrected using the explicit reviewed
reconcile-census.cjs --apply workflow. Existing entry order and all unrelated
entries/reasons remain intact. Three existing reasons are extended; one new
read-only token getter is added. No baseline dump/regeneration or gate bypass. census-reconciliation.json retains individual before/after identities and
lifetime reasoning:
$run: first-zero before last_run_halt/enclosing restoration, saved fields N/A; legitimate CACA0005 reaches it. No writer provenance or ownership proof.$wnd_send_message_inner: pre after caller save/child allocation and before callback frame; event2 before depth/token checks and restoration. Token is current child, not proven modal owner. Scripted abandonment and OFF controls.$win32_dispatch_inner: event3 after teardown/reentry and sentinel adjustment, before current-frame consumption; event4 after previous pump/EIP/ESP restore, before ended reset. Saved ESP denotes consumed frame, not original API entry.export:null_diag_token: synchronous read-only invocation_current getter; copying its signed64-bit representation does not acquire/retain ownership.
Evidence and exact verification
Evidence directory: docs/re-notes/shadow-man-null-arming-identity-repair-20261023/. checks.json records actual argv/env/cwd/exits and stdout/stderr files. Main commands, all exit0:
GIT_NO_LAZY_FETCH=1 python3 docs/re-notes/shadow-man-null-arming-identity-repair-20261023/run-checks.py
node --check test/run.js
node --check tools/shadow-man-null-capture.cjs
node --check tools/shadow-man-null-preload.cjs
node test/test-shadow-man-null-capture.js
node docs/re-notes/shadow-man-null-arming-identity-repair-20261023/old-controls.cjs
node docs/re-notes/shadow-man-null-arming-identity-repair-20261023/check-scoped.cjs
The wrapper uses recorded env overrides to execute the existing donor compiler
in place and retain outputs in this evidence directory. Compiler identities and
footprints are in admission.json. Its installed parser differs from the missing
candidate compiler blob; this is an explicitly identified focused test tool,
not a claim to have qualified the candidate's full vendored compiler provenance. Original902-byte fixture SHA256
f45414f005d33707d05e0a16d9abf6d77f2ddce5aa00ddcddc73551e68e500dd;
actual-path3394-byte fixture SHA256
5b874002a970b97d246408739811bd1802a8c31a6dc2357973579ce67c55f8ab. Both have fixed128-page initial/max8MiB linear memory, no guest/heap allocation. WAT/binaries/source extraction and exact commands are retained for recovery.
repair-tests.json records normal success exit86 and all rejecting argv,
input/submitted-byte hashes, reasons, stdout/stderr/output, limits and explicit
no-continuation checks. Altered custom-section module, substituted bytes,
unbound Module, alternate forms/constructors/streaming, repeated contexts,
wrong PID/runner/hash, conflicts, child and exclusive-output controls exit1;
missing exports exit86 with capture-failed. Immutable bytes/retention failure,
quota/starvation, late/duplicate/wrong action/HWND/context and required-block
failure are also checked. VM execution of the exact runner branch covers
modal resolution, renderer parent climbing, slot fallback, absent dialog and
default-off send. A deliberately late adapter causes the actual ordering
oracle to exit1, retained in late-arm.stderr.
old-candidate-controls.json reuses the review's read-only original fixture and
negative driver with the exact old helper SHA. Eager startup zero and altered
custom-section module both falsely succeed at exit86; each updated-contract
assertion actually exits1. The corrected normal control still succeeds and
the altered-module control now exits1 before compilation/instantiation. These are real rejecting controls, not re-labelled old success receipts.
The native fixture executes the full actual sender body and exact modal sentinel/completion prefix, CACA0005 branch and zero predicate with scripted dependencies. It covers OFF caller restoration,64-round abandonment, teardown mutation ordering, old32-event starvation, and corrected OFF Configuration/OK plus40modal completions before selected-send reset/legitimate-zero capture. Actual full handler / full module / actual canonical export compatibility: NOT EXECUTED. Source export declarations are not an actual module inventory. No guest continuation/writer, performance, persistence, Play or gameplay claim.