Shadow Man callback census correction — source inventory only
Historical source last updated, 2026-10-28. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.
Candidate: 2e6f5984d3f64d7781fb02da09bbb8d564302123. This branch changes
only eight entries in tools/callback-ownership-census.json, adds the focused
tools/shadow-man-callback-census.test.js regression, and retains this review
and its evidence. No production WAT, host JS, scanner, or nonlocal gate changed.
The exact candidate's unchanged production scanner rejected six changed
producers and two unknown producers. The same scanner now accepts 410 source
boundaries: six renewed entries, two additions, and 402 untouched entries. Every entry still says ownership: pending. This closes this source-inventory
blocker; it does not prove callback ownership or accept gameplay.
Eight-producer disposition
| Producer | Source review supporting disposition |
|---|---|
05-alu:$th_ret |
After the existing 16-bit guard, reads guest ESP and its return word once into locals. Kind 2 snapshots that word before assigning EIP from the same local; ESP still advances by 4. The target AST changes from an inline load to the local containing that load. No owner acquisition/release. |
05-alu:$th_ret_imm |
Same load/snapshot/assignment ordering, kind 3; ESP still advances by 4 plus the immediate operand. The return-word writer remains unknown. |
05-alu:$th_pop_run_ret |
Existing stack_run_pop early return and guard remain. The existing fast sr_delta load or gl32 load is captured in a local, snapshotted as kind 4, then assigned to EIP. ESP remains cur+4+imm; no new callback lifetime. |
09b-dispatch:$win32_dispatch |
Preserves resource-pin save/restore, API depth, and existing redirects. The only added operation invalidates diagnostic kind before deferred DLL unload can redirect after the inner handler. Existing EIP targets and ownership boundaries are unchanged. |
09b-dispatch:$win32_dispatch_inner |
Existing CACA0005 completion sets handler_set_eip=1, then EIP=0 to end nested run. The new kind-1 snapshot precedes this write; no ESP pop is introduced. The helper rejects inline-dispatch attribution unless explicit run thunk entry and EIP/thunk equality hold. The enclosing synchronous caller still owns caller-context restoration. All other inventoried targets are unchanged. |
13-exports:$null_transfer_record |
New instance-local, overwriteable diagnostic snapshot. Mode-2 guard, clear-kind-before-rejection, nonzero-target rejection, and explicit kind-1 route guards precede capture. Copies already-read word/address, predicted post-ESP, diagnostic/modal state, validated current invocation token and root; publishes kind last. Writes only diagnostic globals. No run/host callback, EIP/ESP mutation, invocation enter/leave, or retained reference. |
13-exports:$null_transfer_check |
New diagnostic invalidator before the first-zero event. Mode guard; count, synchronous depth, current token, root or post-ESP mismatch clears kind. Writes only diagnostic kind. invocation_current calls the existing validator and reads the token; it can trap on invalid context and does not retain an owner. Matching fields establish bounded correlation, not a writer/lifetime proof. |
13-exports:$run |
Fresh/reentrant entry clears kind and thunk-entry flag. Each nonzero loop boundary clears them again. Explicit thunk dispatch brackets the entry flag. The zero branch checks the record before null_diag_event increments count and calls the synchronous host observer; a returning observer traps. Existing budget/API-depth restoration and auto-pop EIP target remain unchanged. |
cs_pop only updates diagnostic shadow-stack depth. The RET paths retain their
existing branch_end path; this inventory renewal does not certify its entire
transitive execution graph. The snapshot helpers have only the direct
invocation_current call, which validates thread/epoch/root and reads the
current slot. They contain no ownership writes or guest dispatch.
Reentry is explicitly limited: these globals are one instance-local record,
not a stack of retained owners. A nested run clears attribution rather than
preserving an outer record. A new nonzero iteration or deferred unload also
clears it; a changed count/depth/token/root/ESP invalidates it at zero. Fields
other than kind are not attributable when kind=0. Arbitrary host reentry,
uninstrumented transfers, shared-state races and original return-word origin
are not proven safe by matching fields or body hashes. No unresolved writer
or null-transfer path is relabeled proven.
Validation and retained rejection
The corrected production scanner passes. The focused regression executes the actual production parser/scanner on all eight functions and passes 58 negative controls: eight omitted entries, 24 missing/unsupported ownership labels, eight changed bodies with unchanged boundary calls/targets, six changed EIP targets even after digest refresh, ten removed guards, and two mismatched boundary lists. Two development assertions were corrected: parser arrays belong to a separate VM realm, and removing the helper's only invocation call correctly yields a stale producer rather than a changed producer. Their raw failures are retained. The portable final test path passes the same controls.
The existing census mutation suite also passed as supplementary evidence. It internally launches one Node --dump child without forwarding the 320 MiB
heap flag; this was discovered after execution. The supervised aggregate RSS
peaked at 408,760,320 bytes, below 512 MiB, and cleanup completed. That child
heap-configuration deviation is retained and is not a compliant acceptance
receipt. The primary baseline, corrected scanner and focused controls ran
directly with the checked /usr/bin/node --max-old-space-size=320 configuration. No passing group was replayed to conceal this deviation.
Recovery and next gate
Independent source acceptance should precede any later canonical build. This review justifies proposing that build after independent acceptance; it does not authorize or perform it, predict its success, or replace runtime proof.