Q2 open-frame damage diagnostics, 1005

Offline investigation adds bounded passive damage feedback to the actual Q2 observation path in tools/quake2-passive-attack.cjs. Capture and progress now return attackInput.damageFeedback: signed HUD health, screen blend, kick angles and raw pmove flags, plus up to 16 received frames and adjacent-frame health decreases. cause remains explicitly UNKNOWN. Existing first-HP-loss halt, input policy, startup flags, caller pins and native lifetime rules remain.

Evidence: scratch/runs/20261109-quake2-frame-damage-cause-1005/. No runtime, box, Chrome, guest input or retry was performed.

Retained original0940 review

Personally inspected review/frame-edge.png (HP100, open pane) and review/frame-cross-result.png (HP98, nearer open frame). Their SHA256 values match the retained capture records. Examined both binding/motion records, all 14 numbered action records, and frame_step/cross_frame hold records.

Observation Received frame Origin Velocity Raw flags
frame-edge after capture, 09:49:46.101Z 1373 44,-32,24.125 0,0,0 4
cross_frame held, 09:50:18.247Z 1653 45.25,-31.375,24.125 0,0,0 4
frame-cross-result after capture, 09:50:19.196Z 1660 36.75,-2.375,24.125 0,0,0 4

The final command requests ArrowUp for 100ms. Its receipt has one wholly-held physical-high sample and release ACK (started 09:50:18.244Z, release timestamp 09:50:18.343Z). The preceding frame_step requests 150ms and releases ACK. These are physical input and received-state observations, not damage callbacks. Crucially, the two HP screenshots are about 33 seconds apart, and their after snapshots differ by 287 server frames / 28.7 simulated seconds. No HP field was retained at the final action's start. Therefore the HP loss cannot even be localized to that 100ms action: incoming damage while waiting remains possible. Constant sampled Z/velocity/flags neither proves continuous ground contact nor excludes a transient fall, environmental/contact event or incoming attack. The screenshots do not identify a damaging actor. Room entry remains unproved.

Root's original0940 audit was 1790 SHA/size rows and 435 valid gzip streams. The separate offline-orderly/before/controller/stdout.log.gz is truncated; its decompression failure, size and SHA are retained in retained-review.json. It is explicitly excluded from complete-stream claims and was not repaired.

Exact shipped binary support and bounds

The existing canonical EXE SHA256 is 2fa987c9390663e230c5b6479285e7c55511e35f80091bfed96f43062c1c810b. No external source or different Q2 release supplies offsets. Retained exact disassembly and executable-byte assertions establish:

Owner/memory/code checks remain active. Any torn ring rejects damage feedback; inactive/unpublished current frames reject it. Old, mismatched, missing or malformed historical slots are omitted; no decrease is inferred across gaps. Current blend must be finite and in [0,1], kick finite. Health stays signed. No RPC/export, hook or guest-memory write is introduced. The existing production composition already hashes this reader, so its generated provider pin changes automatically without changing shared callers or another game.

These fields retain useful causal inputs, not causality. Blend combines visual effects; kick does not identify a shooter; health decreases are not T_Damage events. Neither attacker, means-of-death, water contents nor contact entity is present in this bounded client reader. Ring wrap and sampling gaps can lose events; repeated equal reads are not atomic history or lifetime proof. Nothing is retrospectively reconstructed for original0940.

Validation and next use

Actual new regression against the old production reader fails because damageFeedback is absent; the changed reader passes. The serialized actual capture/progress reader is tested with authored health100->98, blend/kick, signed health, bounded wrap, gaps, malformed values, current/history races, owner retirement and zero shared-memory writes. Authored data is not original game evidence. Existing passive reader tests pass. Focused real composition normal/held-UNKNOWN/torn/wrong-module/unpublished cases pass, including exact caller drift refusals and unchanged caps. Locked tools/build.sh passes. An initial composition invocation omitted its required output argument and failed before scenarios; its log is retained separately from the passing run.

Next separately authorized runtime can consume the new pinned reader and retain health/feedback before, during and after the first loss. This task authorizes no new runtime or route. Box ID/state/expiry/stop receipt: null, NOT_CREATED. SingleTaskSTOP.