[ConWan30]

Draft. Not for merge without Con's word. Practice lanes only.

What it does. With QORCAP_AUTOCLOSE=1 and QORCAP_LANE=practice (the practice LIVE launcher sets both) on a Call of Duty match, QorCap stops on its own after a confirmed VICTORY or DEFEAT end screen. It waits a tail first (default 12 s) and then stops through the same path as Enter.

What QORCAP_AUTOCLOSE did before. It was read only by the launcher .ps1, which closes the console window 10 s after a clean exit. Neither qorcap.exe nor the sidecar read it, and it was not specific to Madden or CoD. Before this PR, CoD stopped by itself only when the full CoD final posted: the end word plus a readable Qortrola30 row with K/D and owner-row consensus. If the row didn't read, you had to press Enter.

False-trigger guard (sidecar/qorcap_cod_endstop.py). It watches reads the CoD frame loop already makes. It doesn't read pixels itself and posts nothing to /events.

• A frame counts only when CodEndScreenReader returns a calibrated, confirmed victory or defeat.
• Any other frame ends the run, so a one-frame flicker cannot confirm. That covers killcam, gameplay, lobby, a mid-match Tab board without the word, a blurred or cut word, and the other word.
• Timestamps must be strictly later, with no gap over 3 s (the reader's own agreement gap).
• If Qortrola30's end row was read on at least one frame of the run, the bar is 8 reads over at least 3 s (QORCAP_AUTOCLOSE_FRAMES). If the row was not read, the bar is twice that: 16 reads over at least 6 s.
• Arming: the run must start at least 120 s after the first frame (QORCAP_AUTOCLOSE_ARM_S). A leftover end screen in the lobby cannot stop a new match.
• Once confirmed, the result stays fixed for the rest of the match.

Stop. After confirmation, capture runs QORCAP_AUTOCLOSE_TAIL_S (default 12, clamped 10..15). Then the main loop stops the same way as Enter: /shutdown, seal, countersign, bundle, receipt.

• The stop reason is an existing value: final if a CoD final posted, otherwise post_final_tail. No new value goes into the receipt.
• Enter/Q and capture errors still stop at once.
• The tail only adds time. An older final stop that would fire earlier waits for it.
• The live tap keeps sending through the tail. It ends the way it does today: QorCap stops sending, and QorStream's UDP idle timer ends the ingest. There is no separate end signal.

Info only.

• One stop note: line goes into qorcap.log. /shutdown accepts an optional printable-ASCII note of at most 240 bytes; it is logged and not stored. The line contains none of the needles stop_reason_from_log looks for.
• cod_end_stop.json is written beside the bundle: result, reader, run count and span, frame seq, capture t_ms, config. It follows the same rule as post_final_tail.json: not bundle.json, not the receipt, not read by verification.
• I did not use cod_session.json, because attach_session copies its sync tags onto bundle.json.

Files. sidecar/qorcap_cod_endstop.py (new), sidecar/qorcap_cod.py (+12: optional read hook; an exception in the hook is swallowed), sidecar/qorcap_match.py (wiring), src/api.rs (/shutdown note logged), tests, tools/assert_pytest_skips.py (new file in both suites; box-only file excluded; floors re-collected to 905 linux / 887 windows against qortime 0204e6d).
Untouched: receipt.v0, contracts/, schema/, the capture encode, the recording path. REAL MATCH stays on v0.3.38. The keyframe fix (Fix 2) stays on hold.

Tests (local Linux).

• test_qorcap_cod_endstop.py, 25 tests:
  • Negatives: killcam, Tab, flicker, word switch, gap, repeated timestamp, pre-arm, another player's row.
  • Positives: VICTORY and DEFEAT, with and without the row.
  • Tail timing with an injected clock, the stop-reason wiring, the shutdown body, the log note and sidecar file, the runner hook.
  • Committed 720p fixtures: VICTORY splash then board confirms; mid-match and transition frames don't; board alternating with gameplay doesn't.
• test_qorcap_cod_endstop_box.py runs only on the build box (excluded from CI) over 370 earlier sampled frames. No mid-match sample read as an end word; each of the 18 end-word samples confirmed its own word when held.
• Linux sidecar suite: 894 passed, 1 allowlisted skip. The 10 failures are all test_qorcap_submit.py because fastapi isn't installed on the box; CI installs it. Skip gate passes.
• cargo test: lib 206, main 1, contracts 2, haptic_replay 1, redteam 2, all pass. No new clippy warnings.

Not measured. I have no committed killcam or "Final Killcam" frames, so the killcam negative is modeled ("no confirmed word"), not measured on a real killcam. Real-match recall at the sidecar's sample rate (about 2–4 Hz) is unmeasured; a slow PC could take longer to reach 8 or 16 reads. In the sampled frames, "BEST PLAY" follows the end header, and a 12 s tail may end during it.

CI artifact. The CI on main uploads qorcap-<sha7>-windows-x64 for this PR, which contains #150 plus this change.