Shadow Man zero-transfer independent source acceptance, 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.

ACCEPT_SCOPED_OFFLINE for candidate 2e6f5984d3f64d7781fb02da09bbb8d564302123, whose exact parent is b41a7c5f4b752d1f0bdab1141a11d4fed4dc6f9a. No source correctness counterexample was found within the selected, single-instance diagnostic path. This accepts the source instrumentation and offline controls, not a compiled module, native execution, original capture, ordinary callback completion, or game behavior. Original 61e1ea2d remains REJECTED; its immediate zero writer remains UNPROVEN. The preexisting original modal differs from the synthetic callback-created modal; zero state alone proves no normal completion.

Actual source control flow reviewed

Boundary Source finding and acceptance limit
CACA0005 writer 09b-dispatch.wat:1335–1343 records after handler_set_eip=1 but before eip=0. The marker is the metadata actually loaded by the dispatcher; the thunk is computed from the dispatched index, rather than inferred from its name. The helper requires explicit run thunk-entry mode and architectural pre-EIP equal to that actual thunk.
Actual versus named thunk Both named continuation values are retained. A CACA0005 marker at an unexpected address remains native-caca0005 with syncReturnThunkMatches:false; this identifies the executed source assignment, not an authenticated intended continuation. No ordinary-return inference is licensed.
RET / RET imm Both read the existing stack word once into a local, snapshot it before EIP/ESP mutation, then perform the original mutations, cs_pop, and branch_end. No extra guest stack dereference is introduced. RET imm uses the actual operand for predicted post-ESP.
Fused POP/RET Record follows successful stack_run_pop and uses post-POP ESP as the RET address. Actual source commits POP ESP, returns early on its reported fault, and chooses the existing fast translated load or gl32 slow read. The record describes the terminating RET, not the preceding POPs.
Return to the run desk Actual cs_pop only updates its diagnostic depth. branch_end delegates to branch_end_at; armed range tracing recomputes both chain and tier guards, causing a desk return before following another threaded block. This closes the gap left by the evaluator's no-op tail stub for this selected adapter.
Four record calls Exactly CACA0005, th_ret, th_ret_imm, and th_pop_run_ret; all are before the corresponding EIP write. The sole check call is immediately before the existing first-zero diagnostic event in $run. Complete locations are in the source receipt.
Fresh/reentrant run Entry clears kind and thunk-entry flag. The zero desk validates before the next nonzero iteration clears them. A new invocation cannot inherit prior attribution; an outer invocation may conservatively lose attribution after nested work, which yields unknown rather than a fabricated writer.
Explicit thunk dispatch $run sets the entry flag immediately around win32_dispatch, clears it afterwards, and branches back to the zero desk. CACA0005 sets handler_set_eip, preventing the normal auto-return pop from replacing its assignment.
Inline dispatch Nonzero desk clears the flag; inline CALL/JMP dispatch has no arming site. Actual inline callers can overwrite EIP after the handler, so these cases remain unknown, even if architectural EIP happens to equal the thunk.
Deferred unload The dispatch wrapper first completes resource unpinning; those helpers release pins, not guest control. If release schedules a drain, it clears kind before any deferred EIP selection/unload. This conservative invalidation can discard a real CACA assignment but cannot falsely name it.
Other writers CACA0004, auto-return, indirect CALL/JMP-to-zero, Win16/far/tier paths, and host-supplied zero receive no fabricated attribution. Run entry/nonzero iteration reset, plus the selected trace guards, bound history to the immediate desk trip.

The record/check match saved i64 token and root, synchronous depth, diagnostic count, and predicted ESP. The zero event increments count once; serialization independently requires that adjacency, zero EIP and matching ESP and depth. Saved modal fields belong to the pre-transfer snapshot; live modal fields remain in the ordinary boundary envelope. No attempt is made to certify all modal-frame history. preEip is the architectural block EIP, not a promise of the exact RET instruction VA; the site and threaded cursor provide context.

Actual CLI and observer contract

The actual CLI selected-send hook is unchanged. The selected adapter binds the instance object, PID, exact runner, dlg-cmd, command 3, action sequence, live resolved dialog and batch. Earlier Config/OK actions run before arming, with bounded retained history. Selection does not reuse synthetic batch/HWND values.

The actual preload checks runner realpath/hash, named module bytes/hash, single compile/instance identity, no start section, no alternate constructor or streaming path, and main thread. Its retain function calls writeSync then fsyncSync. arm retains the pre-arm receipt synchronously before trace configuration, native mode 2, and the actual send. A retention failure stops before send; unsupported protocol version stops before arming. These source ordering facts were inspected; this review did not execute the real CLI or intercept a real OS fsync/exit. A future watchdog must independently bound synchronous execution; the metadata does not enforce that watchdog.

Executed offline controls

Results: 24/24 candidate checks PASS, 7/7 independent lifecycle checks PASS, exact-source audit PASS, all exit 0 with empty remaining process groups. See candidate receipt (historical artifact: shadow-man-zero-transfer-independent-review-20261023/candidate-offline.receipt.json), independent receipt (historical artifact: shadow-man-zero-transfer-independent-review-20261023/independent-probe.receipt.json), and source receipt (historical artifact: shadow-man-zero-transfer-independent-review-20261023/source-review.receipt.json). The copied candidate test changes exactly one ROOT path to read the standing candidate in place and write outputs only beside the owned test; the input hash and replacement are retained in test-adaptation.json (historical artifact: shadow-man-zero-transfer-independent-review-20261023/test-adaptation.json). No donor files are written.

The independent probe uses its own small parser/evaluator for the actual run entry clears, zero-desk prefix, explicit thunk arm/call/clear/auto-return path, CACA branch, record/check helpers and deferred invalidation. It checks no stale writer across fresh/reentrant entry and nonzero iteration; unknown inline dispatch even with matching EIP; token/root/depth/count/ESP invalidation; and positive immediate CACA preservation. Three intentionally defective in-memory mutations independently remove run clear, iteration clear, or inline gating. Each then triggers the rejecting false writer inherited assertion. They are rejecting controls, not defects found in the candidate; no production fix was made. The probe models thunk bounds/metadata, invocation reads and unexecuted run service helpers and does not run the whole interpreter.

The 24 candidate tests faithfully exercise the changed forms they select and the real JS serializer. Their full-run claim is narrower: run/deferred checks extract particular resets rather than executing whole control flow. Fused POP setup is stubbed, fast mapping is prearranged, cs_pop/branch_end are no-ops, invocation ownership is stubbed, and pre-arm retention is a mock rather than an actual filesystem/CLI. Thus they cannot prove canonical compiler behavior, real fused POP address selection, memory traps, native invocation validation, reentrant execution, actual CLI process binding or OS86 termination. The source walk above checks surrounding implementation but does not upgrade these models to native evidence.

Source/census/ABI impact

Exactly +96 instance-global value bytes: 20 mutable i32 plus two mutable i64 globals, with engine storage overhead additional. Existing field accessor protocol version is 1 at field 9; mode 2 opts in, fields 10–28 expose the record. No new named import/export or linear region is claimed or found in the source diff. Function/global indices, code bytes, and resulting module identity will change. This source census is not a compiled ABI inventory. Producer hashes, compiler files, census fixtures, and acceptance oracle were not refreshed or repaired. Historical module 8fac39936b6df2418ad86cab118cef19a65a727f649fbc14c412b37818904686 is not evidence for this candidate and was not copied or instantiated.

Exact remaining qualification, future coordinator assignment only

  1. Under a fresh compatible build allocation, pin this candidate's complete src/main.watx include closure and exact compiler path/blob/full hash plus byte-exact excerpt recipe. Produce the corrected static ABI-helper receipt including both storage regions and named spans (including DIRECT_WINDOW), with omission/change/duplicate rejection controls. Do not inherit historical producer acceptance or a storage-only fingerprint. Resolve required producer/census gates under a separately assigned change; this review has not authorized skipping them.
  2. In that complete qualified checkout, execute the actual canonical command bash tools/build.sh and retain every mandatory gate result, new module SHA-256/size, ABI census and reproducible source/compiler binding. Never use wat2wasm, a helper-only build, or the historical module as substitution.
  3. Native acceptance must use that exact complete module, real ownership initialization and a future focused fixture that invokes actual run, dispatcher and decoded guest RET paths. Cover explicit CACA0005 (including actual/named mismatch), C3, C2 imm16, fused POP/RET fast and slow routes, nonzero/unknown/inline/deferred/reentrant/stale cases and modes 0/1/2. Require pre-mutation fields and real token/root/depth/count/ESP adjacency; use the unchanged capture adapter and required-order rejection. Assert no false attribution on all unknown controls. Returning observer must trap; process capture must fsync pre-arm before native arm/send and terminate OS86 with all groups reaped under an external watchdog. A native fixture for these new writer paths does not yet exist in this review; creating and qualifying it is a concrete remaining dependency, not an implied PASS.
  4. Existing executable regression entry points to retain are node test/test-shadow-man-null-capture.cjs, node test/shadow-man-null-actual-paths.cjs, node test/shadow-man-null-repair.cjs, and node test/shadow-man-null-fresh-selection.cjs, each under its admitted guard. The first two assemble extracted fixtures and are not substitutes for the complete-module gate. Their declarations must first be reconciled with new helper dependencies: for example test-shadow-man-null-capture.cjs injects the new header but its handwritten globals omit reg_base, ip and the named continuation globals. This is an inspected fixture dependency, not an executed compiler failure; this task neither ran nor fixed it.
  5. Only after those gates, authorize one original actual-CLI observation with the new qualified module, exact unchanged runner and this candidate's preload/capture closure. Use SHADOW_NULL_MODE=revised-first-zero, SHADOW_NULL_ZERO_TRANSFER=1, SHADOW_NULL_BLOCK_POLICY=required, explicit newly built --wasm, pinned SHA, PID/runner binding and external watchdog. Observe fresh Configuration, OK, new launcher and counted final ordinary Beenden; retain actual argv/action/instance/module/fsync/native writer/terminal evidence. A required-block/order failure remains rejection even with a known writer; never resume past zero. This can distinguish the historical fork but cannot retroactively prove the old session's writer, settings, Play, persistence or gameplay.

The canonical and existing regression commands above are executable future entry points, not commands run here. The new complete-module writer fixture and fresh original action selection require a new bounded assignment; there is no already-qualified native acceptance command for this candidate to reuse.