feat(pipeline): widen the reasoning window to see placements (gate item 6)

[AlexSKuznetsov]

What

Placement detection reads change — the reasoner contrasts widely-separated frames and sees "holding" → "put down, walked away". A tight window shows only continuous motion toward the surface, so the transition is invisible. Spike 5b measured this: a ±5s window took detection 2/2 → 0/2 on real clips. The shipped reason_window_seconds was 6s — tighter still.

Knob	Was	Now	Why
reason_window_seconds	6	20	Straddle the placement transition; stays under the 30s evidence ring
reason_max_frames	4	8	Over a 20s window, 4 frames sit ~7s apart and fall either side of the transition
event_cooldown_seconds	20	25	Wide windows now overlap, so one placement is seen by several — cooldown must cover a window span

Two model_validator invariants lock the coupling (window ≤ ring retention, cooldown ≥ window), with new test_config.py.

Local validation on LFM2.5-VL-3B (proxy for Cosmos)

Ran on spike 5b's own clips (clips/recordings/VID_20260819_1207{27,02}_{hi,win}.mp4; _hi = full 720p, _win = ±5s 720p), sampling frames endpoint-inclusive exactly as cosmos does:

object   arm      K=4   K=5   K=6   K=7   K=8
keys     narrow   NO    NO    NO    NO    NO
keys     full     NO    NO    NO    NO   PLACED
wallet   narrow   NO    NO    NO    NO    NO
wallet   full     NO   PLACED PLACED PLACED PLACED

• The narrow ±5s window never detects the placement — at any frame count, either object. Window width is the binding constraint (reproduces 0/2), confirming the core fix.
• A wide window detects it, but only if sampled densely enough: the large wallet from 5 frames, the small/thin keyring (the demo's hard case) only at 8. At production's reason_max_frames=4, even the wide clip missed both — so widening alone is not sufficient, which is why reason_max_frames is bumped too.

Caveats: LFM is a proxy for Cosmos; the absolute frame floor (8) is model-specific, so reason_max_frames is a knob to re-tune against real Cosmos (watching the 8192-token context and latency). The direction — narrow never works, density matters for small objects — is robust.

Notes

Cosmos already endpoint-subsamples the window (np.linspace), so widening the window costs tokens only through the frame-count bump, not the width. A wider window also yields a richer evidence clip (hold → place → walk away). Full suite 178 passed, ruff and pyright clean.

🤖 Generated with Claude Code