TR2 weapon consumption after Action — 2026-11-08

The closed 1810 original proves native Control → successful keyboard poll → Action EBX 0x40. It does not show whether that mask reached the weapon update, which weapon state ran, whether an arm reached its firing frame, or whether target/aim/ammo gates admitted a shot. The observer removed itself at that Action pair, before those reads. Repeating a blind hold cannot distinguish these cases. This change adds an opt-in, bounded, read-only continuation through one linked weapon update, supported by offline exact-byte causal fixtures. No original, browser or remote box was started for this task.

Evidence directory (R): scratch/runs/20261108-tr2-weapon-consumption-1832/. Starting Git SHA, relevant owned binary diff, small owned/helper patches, provider hashes, test logs and static disassembly are retained there. Git is the source archive. Existing media and module are referenced by path/hash; there are no source-tree archives or media copies.

What the retained original actually established

Startup/Control report retains the sole 2000 ms Control interval at 18:18:40.817–18:18:42.817 UTC. At boundary 22,368, PC 0x447206 returned HRESULT 0, DIK1d/9d were [128,128], VK17/162 were down and VK163 clear. Slot8 was configured156/default157/conflict0. Its paired successor 0x44d3d4 recorded keyResult1, OR executed, EBX40 and Action true. Two records were retained, qualified and removed. No downstream weapon snapshot was taken. Lara was upright with drawn pistols before the hold; the dog was close and health about three-fifths afterward. No reviewed flash, hit, kill or effective aim was established. These are retained observations, not evidence that an input binding repair is needed.

The earlier input diagnosis, Action observer and coverage repair remain applicable. Their historical manifests and results are unchanged.

Exact original dependency map

All addresses below are original VAs in test/binaries/candidates/tomb-raider-2-demo/game/TOMB2.EXE, SHA-256 0811efa84bfab8c458b9e29976088152db4c14601a99ca4910c75a2d009233aa. Numeric states and branches are established by this binary; descriptive weapon names follow its familiar paired-arm behavior, not imported symbols.

Stage Exact bytes/address behavior What this distinguishes
Publish Action 44dd55 loads the final local input mask; 44dd63 stores EDX to 51b178. The alternate early exit stores EBX at 44d49f. Returns are 41427f, 411fd8, 4220a7. EBX40 at slot8 precedes final mask publication. Later input assembly and demo/control code must not be assumed to preserve it.
Weapon update 42e640; exact call returns 4275ac, 4275c2, 431757, 432111. Decrements arm flash counters, reads Lara item 52173c, HP at item+22. Nonpositive HP forces status0. Whether the linked control flow actually reaches a weapon call and with what health/body animation.
Before dispatch Alive path 42e949; status4 path 42ea88. Input20, requested type differing from current, and water/depth conditions can change status to3 before dispatch. Water0 permits the ready route; water4 also compares depth 52168c to negative weapon-info height. Draw/holster/request/water gates are separate from Action40.
Status dispatch 42e68d: signed word 521662, range0..5, table 42eb28: destinations 42e876,42e8e7,42e6a4,42e719,42e775,42e76b. Status4 reaches 42e775. Drawn-looking pixels alone do not establish numeric ready status at the relevant call.
Ready Action/ammo 42e7ba tests input40; clear goes directly to type dispatch. Set selects ammo through 42eb80; 42e803 requires signed ammo >0. Failure 42e807 clamps ammo0, plays empty sound and chooses requested weapon, then returns. Ammo can block before arm animation. Type1 uses 52171c; types2/3 use 521720/521724.
Type dispatch 42e83c, tables 42eba4/42eb98: types1–3 call 42d330 (return 42e860); types4–7 call 42bc70. Paired-arm instrumentation cannot infer equivalent behavior for other weapon types.
Target/aim 42d345 tests input40. Clear drops target 5216e0; set calls 42ed30. A null target causes acquisition 42eed0. Then each arm is interpolated through 42f1a0, before 42d452 calls arm processing 42d4c0. Target existence, lock and arm angles can differ even while Action is set.
Arm gate Right: 42d4df..42d500; left: 42d688..42d698. Lock permits raise/fire branch. Without lock, Action+null target also permits it; an existing unlocked target enters lowering branch. A target is not a universal prerequisite to firing; an existing unlocked target can suppress it.
Firing frame Right 42d54a..42d56a; left 42d6d7..42d6f4. Frames0–3 raise; frame4 additionally requires Action40. Frames24 onward wait for weapon-info cooldown before resetting to4. Action consumed for raising/cooldown is not yet a shot.
Shot Right call 42d595 / return 42d59a; left 42d71f / return 42d724; callee 42f270, arguments type,target,Lara,angle-pair. Exact entry/return separates attempted shot from arm motion.
Ammo inside shot Type1/default path 42f2f9 resets 52171c to1000, then 42f308..42f315 checks and decrements it. Types2–5 select other counters, with cheat-byte 51fd96 reset behavior. Nonpositive ammo returns0 at 42f5c1. The pistol reset does not bypass the earlier ready-state ammo gate.
Return consequence 42f505 calls target damage 42f5e0, then returns1. That primitive subtracts HP at target+22 (42f605), with a separate kill-counter threshold. Alternate collision/miss routes return FFFFFFFF. Both nonzero returns trigger arm flash/sound; zero skips those writes. A nonzero shot result is not equivalent to target damage or kill. Raw result and same-target HP must be interpreted with the exact path.

42ed30 clears both locks/target angles when target is null. With a target it computes relative angles (42f0f0, 401250), calls LOS 415a60, and tests weapon-info bounds. The post-LOS return is 42edec; LOS0 clears both locks through 42eea3. The primary window sets both locks; outside it, 42ee40 applies separate retention windows per arm. Type1's primary yaw/pitch bounds are ±10920; one yaw retention range extends to30940, the other to−30940. Thus authored yaw30000/pitch0 retains one arm and clears the other. This explains a possible mechanism, not the cause of the 1810 result: its actual angles, LOS, target, ammo and animation state were not retained.

State addresses: type/request/last 521664/521666/521668, water 52166c, target yaw/pitch 5216e4/5216e6. Arm structures are 5216f8 and 52170a, stride18: frame+4, lock+8, yaw+10, pitch+12, flash+16. Item animation/frame are +14/+16, body state+0e and yaw+42. Ammo snapshot covers 52171c..521734.

Observer contract and tests

tools/tr2-action-observer.cjs defaults to the existing Action-only behavior. Explicit weapon:true continues from the first paired poll/Action. It matches the saved input caller and exact stack displacement +6c, records the stored input, requires the same full mask at weapon entry, and matches the four known weapon callers. Exact fixture calls establish nested stack displacements −24 (paired arms), −40 (arm processing), −80 (shot), and their return addresses. It records status/type/request/water/depth, target/HP/angles, each arm's frame/lock/angles/flash, all seven ammo counters and Lara HP/animation/frame. Optional observed post-LOS/gate boundaries retain raw LOS result and state; shot returns retain raw EAX and the actual argument target's HP. It stops at the first linked weapon return. It does not repeat input updates looking for a better result; a superseding input update ends unqualified.

consumptionQualified means a successful Control/Action pair reached a completed, linked weapon update. It does not mean status4, shot, damage, kill or effective aim. The raw gate/shot records make those distinctions. Clear Action can still produce a complete diagnostic trace but cannot qualify Control consumption. Out-of-hold returns do not qualify the native correlation. Internal WAT reentry and other threads remain unobserved. No row is absence proof; a missing required call link, changed input, unmapped read or incomplete shot fails closed. A complete status0 trace can explain the observed branch without claiming readiness or firing.

The original 2500 ms installation deadline and 32-record cap remain. Activation follows native keyDown without resetting the deadline. Independent native release remains in the unchanged hold helper. The facade debits actual retired blocks, including fold overshoot, preserves aggregate reporting and forwards the remaining caller budget after removal/expiry. No extra guest budget, setter, translator/materialization, tracing flag, clock, seed, binding/backend change or guest write is added. A Worker timer cannot preempt an in-progress WAT call; external lifecycle bounds remain necessary.

tools/tr2-weapon-consumption.test.cjs passes 126 checks, including exact original pin comparisons; immutable instruction slices on production CPU d8facd0658b7f158c6e60bef2de0936471f438f8bf41e9240fd75635f1d319dc; status/type dispatch, positive/negative target/lock/Action/frame/ammo branches, post-LOS primary/retention windows, exact call-stack links, input publication, and damage subtraction/kill threshold. Registers, stack, geometry/LOS results and data are authored fixture inputs; ballistics stop before RNG/dependencies. No PE loader or game loop runs. Observer state fixtures cover both arms, zero/clear states, wrong caller/code/input/stack, missing aim, unmapped target, pending return, supersession, expiry, overflow, removal and unchanged guest state. Serialized opt-in orchestration tests activation after native keyDown, in-interval correlation and independent release despite stalled arm response. The unchanged 32 Action-observer checks cover retirement/deadline contracts; production direct/RPC input path and helper-closure tests also pass. The locked full build passes with identical d8fac module bytes.

Prepared selection and next assignment

R/prepare.py relocates the normal 615-provider / 29-helper closure from 1750 SHA-256 d47d09a5520ce6bc4d2321c90f9e4b45446f5bed182fd3fd3248912bcd4bd72d. Only tools/tr2-action-observer.cjs changes among non-helper providers; the other585 remain byte-for-byte selected. The relocated browser helper passes weaponObserver:true only with explicit actionObserver:true, preserving the single ordinary Control ≤2000 ms admission. Relocation also applies the current owner disk-only 1.5GB rule. All615 source providers resolve by hash, all29 physical helper targets validate, and the actual driver preflight exits before Chrome. Existing85 retained coverage artifacts are independently hash-checked and reused. Historical614 and prior615 selections are immutable.

Next is one separately assigned original, using this explicit selection, fresh owning context/module/HTTP authority, reviewed gameplay and the same single independently released Control interval with both observer flags. Inspect the linked status/gate/frame/target/lock/ammo/shot records before assigning another action. No original or remote allocation is authorized or performed by this preparation task.

Fresh owning WebGL2's Intel string versus SystemInfo SwiftShader/SwANGLE is still unresolved; no physical-GPU claim. Full27, parity, audio and FPS remain open. Task box ID/state/expiry/stop receipt: N/A (none created). No public deploy, old resume/delete, archive removal or subworker. Single-task worker stops after tested commit and immediate authenticated publication.