TR2 bounded passive target and damage observation — 2026-11-08

The existing observer now retains the shot argument's native target pointer, object-number discriminator and signed health at entry and exact return. It also watches the synchronous damage call inside that invocation, storing its evidence in the existing shot-return row. No extra records, guest calls or writes, observation flags, runtime/input/config/clock changes were introduced. The 2500ms/32-record bounds and actual-retired debit/remainder forwarding are unchanged. Code: tools/tr2-action-observer.cjs; focused controls: node tools/tr2-target-damage.test.cjs.

Independently inspected native boundaries

Installed original: test/binaries/candidates/tomb-raider-2-demo/game/TOMB2.EXE, SHA256 0811efa84bfab8c458b9e29976088152db4c14601a99ca4910c75a2d009233aa. The run's input-path-hashes.json gives independently read size/CRC/SHA and existing CPU module paths. No media/module copies were made.

Native acquisition at 42eed0 walks short item indices, resolves them through the item-array pointer at 527270 with stride72 (42ef5d..42ef6f), rejects nonpositive signed health at offset22, tests geometry/LOS and selects the pointer stored at 5216e0. This is a native ITEM pointer, not a renderer ID. Offset0c is the discriminator tested against35/36 by the damage routine. Pointer and discriminator are evidence, not a lifetime-generation handle. The observer captures target presence; it does not newly qualify acquisition geometry, LOS, lock effectiveness, object names or successful aim.

Existing causal links remain poll → Action → published input → weapon caller and stack → paired arms → shot 42f270, with exact shot returns 42d59a or 42d724 and ESP entry+4. The shot prologue reserves254 bytes and saves four registers. At 42f4f0..42f505 the original pushes damage, position and the shot target argument, then calls 42f5e0. Its entry ESP must be shot-entry ESP−274, saved caller 42f50a, target equal to the nonnull shot argument. Return must be that exact caller and entry ESP+4. New literal instruction pins are checked against the installed original by the focused test.

42f605 subtracts the low damage word from signed target health at+22, then sets a flag at+46. The observer records damage-entry health, amount, return health, exact boundary numbers and signed16 expected result. Positive damage requires that linked call to return, unchanged object discriminator across shot/damage boundaries, a matching subtraction, positive amount and decreasing health. Shot-wide health delta is retained separately: delta without the linked damage boundary is not qualified damage. Raw EAX1/ffffffff, ammo decrement and target presence alone establish none of this.

Missing damage entry/return boundaries are unknown, including internal WAT reentry, other threads, damage after this return, or later projectile paths. A seen entry without its exact return refuses shot completion. Same-pointer, same-discriminator recycling cannot be detected. Crossing zero is not a kill claim. No ordinary original was launched to qualify the new fields.

Retained original and new checks

Run: scratch/runs/20261108-tr2-target-damage-observer-2131/. The actual retained2116 observer stream was decompressed and independently checked against its recorded original size/SHA, including gzip CRC. Its26 rows still show shot-entry24/return25, type1, target0, EAXffffffff, ammo1000→999, Lara HP1000 and pendingWeapon true/pendingShot false. It proves an attempted shot only. The historical result is unchanged; new fields are not backfilled.

Focused real-x86 controls use existing accepted CPU d8fac by path/hash and immutable original shot-call/damage/epilogue slices. Stacks/data are authored; the blood dependency is explicitly an authored RET. No geometry/RNG/AI/game loop is executed. Controls cover positive/zero/negative damage, health crossing zero, stale object discriminator, target0, native exact−1/0 returns, wrong damage target/caller/stack, unlinked or pending returns, wrong shot return stack/caller, stale publication generation, wrong owner, and delta without damage. Whole512MB hashes check active observation boundaries for zero writes; record overflow and deadline removal remain bounded. The locked flock /tmp/wine-build.lock bash tools/build.sh passed. Unchanged historical proof suites were not repeated.

Evidence inventories every regular owned-run file, including extensionless files if any, with independently reread gzip CRC and original SHA/size. Git SHA plus small owned patches preserve source recovery; no tree archive or peer WIP is captured. Preparation exception: the first disassembly/head streams were initially written through a subprocess file descriptor that bypassed the gzip wrapper; their unique bytes were losslessly compressed afterward. All focused/build/watch streams used pipe-to-gzip from their first stored byte. No old runs were deleted. Prepublication disk samples were above1.5GB. The final shared-tree sample fell to1,499,463,680 bytes during concurrent writes; the owned run holds about26KB. No unauthorized cleanup was attempted. No continuous floor claim or RAM gate is made.

Box NONE, state NOT_CREATED, expiry NULL, stop receipt N/A. Offline only: no Chrome/game/original/remote/native input/new workers/public deployment. Full27, parity, audio, FPS and renderer limits remain open. Next is a separately authorized ordinary original using the new observer identity, current target evidence and these exact bounded links; this task authorizes no such run.