Quake II source lifetime revocation, 2026-11-08

The default-off owning CPU observer now revokes acquired-source lifetime records at actual native release/allocation entries and existing reset/import/export/error boundaries. The historical acquisition SHA remains a point observation. It never certifies current or later bytes, even when both the address and bytes recur.

Evidence: scratch/runs/20261108-quake2-source-lifetime-2335/. Offline only, single worker, no browser/game/box. Box NONE / NOT_CREATED, expiry NULL, stop receipt N/A. No Q3, SuperTux, backend, clock, timeout, game, input or instruction-stepping changes.

Implementation

tools/quake2-owning-load-witness.cjs adds bounded sourceLifetime metadata: at most 32 sources (bounded by the existing record cap) and 32 revocation events. Each source references its historical acquisition record and transitions from ACQUIRED_POINT to sticky REVOKED. ACQUIRED_POINT itself asserts no allocation liveness. currentIdentity stays UNKNOWN and laterBytesCertified stays false. Snapshot arrays/objects are copied, including event membership lists.

The original installed EXE supplies separate pins for FS_FreeFile 0x416ef0, Z_Free 0x414890, Z_TagMalloc 0x414950 and CRT free/malloc 0x432c10/0x432c80. At actual owning block entry the observer checks CPU PC/ESP and original bytes. Release revokes matching payload/header addresses before any native free store; allocation conservatively revokes every outstanding source before allocator stores. Unverified substitute bodies revoke conservatively without claiming native release. Release entry means attempted release, never successful free. A bad-header error must also revoke. No caller completion/generation flags are used. Historical native load returns remain separate from source lifetime.

Original path exercised

tools/quake2-source-lifetime.test.cjs reuses the hash-pinned retained2205 driver and its verified installed EXE/PAK/BSP inputs by path. It runs the shipped CPU, unmodified NodeWorker, actual command queue, local observation attachment and resolved PE imports. Native/CPU/module/media copies are not staged in evidence.

Unlike2320's retaining free-counter stub, this fixture executes the original:

The PE loader resolves IATs. Fixture startup creates a real heap through HeapCreate, initializes lock9 through InitializeCriticalSection, and supplies empty zone/small-block-list state and threshold0. These are explicitly authored CRT startup state, not an original process-startup qualification. Allocation occurs before CM_LoadMap entry. Fixture file acquisition copies the original BSP into that actual allocation and returns its pointer/length. Cvar_Get and flushmap remain authored native fixture boundaries. OS file I/O and complete native CRT startup are not qualified. No allocator/free routine is replaced.

Original-body assertions cover the file release, zone wrappers/allocator/free, CRT wrappers/large-block allocator, small-block lookup and lock/unlock bodies. The source is allocated at 0x7ee10014 in these controls, distinct from input0xc00000. Native MD4/checksum0xbfd75753 and all35 cmodels match independent retained2205 assertions in the final reuse controls. Fixture startup and file read have no host register edits between native loader entry and return/failure.

New controls and parity

Seven distinct off/on pairs (14 Workers), all terminated and awaited:

Every pair compares final architectural registers/yield state, lazy flags and the complete direct guest window. The final three controls additionally compare the actual translated high allocation's header+payload SHA. Required locked build passes. Existing accepted2244/2258/2205/2320 checks are reused by PATHHASH; none were identically rerun. The initial allocation-inside-loader diagnostic failed closed on a real host import before acquisition. The fixture was changed to allocate before loader entry. An initial pending-reset test used an incorrect MD4 PC range and failed; the corrected original MD4 range passes. These failures are retained, not acceptance evidence.

Native HeapFree produces actual host-import logging boundaries. Those retain the existing conservative sticky invalidation; these original allocator runs issue no native load-return continuity claim. This task closes the concrete retaining-stub lifetime experiment gap and adds explicit lifetime revocation; it does not establish all-writer/exclusive identity, complete instruction coverage, native-return-to-later-use identity, or edict ABA.

Recovery and next prerequisite

Git SHA plus small owned binary patches recover sources. All regular run files have original/stored SHA256 and sizes, GNU gzip CRC and independent original reread verification with an exact-set inventory. Logs are gzip from first stored byte. The first boardwatch gzip was killed before EOF; its incomplete stored bytes are retained inside a valid gzip, explicitly marked as incomplete rather than claimed as a complete log. The replacement watcher has an independent 3600s timeout and is drained with actual same-shell waits and gzip EOF. /home free samples exceed1,500,000,000 bytes; no continuous-floor claim, RAM gate, archives/media/module copies, peer/old deletion or agents.

Before verified guard/onward-door gameplay: establish coverage of actual writers and exclusive ownership from acquisition/conversion through server startup and later table/headnode use, with missed reload/reset rejection, plus a lifecycle-complete edict ABA witness. Native completion and source reuse tests alone do not meet that prerequisite. Single task STOP.