Adaptive loop-detector windows

LoopDetector flags three loop shapes — byte-equal repeats, coordinate-drift, and frozen-state. The runner responds with a soft-nudge at three repeats and a hard-terminate at eight. Those fixed thresholds work for the median case but hurt two patterns:

• Pagination / drilldown. The brain emits five identical click("Next") actions; the page advances each time. The legacy is_repeat_loop fires (identical params) even though the run is making progress.
• Stuck loops with micro-changing pixels. A captcha or animated spinner re-renders enough to defeat is_state_loop's frame-hash check, and the brain takes seven dead steps before the hard window trips.

This module adapts the comparison window per call by recent action diversity + observed state progress.

Tracking issue: #298.

Surfaces

Method	Returns
LoopDetector.pattern_diversity(window)	float in [0, 1] — unique action signatures / window. Click-like actions bucket coords by ~click_tol_px so micro-drift collapses to one signature.
LoopDetector.state_progressed(window)	bool — public mirror of the URL/frame-hash check already used by is_repeat_loop.
LoopDetector.adaptive_window(base, max_extension=2, floor=2)	int — base widened on diverse / progressing windows, tightened on low-diversity / frozen-state windows. Floor and max-extension clamps applied.
LoopDetector.is_any_loop_adaptive(base, max_extension=2)	bool — is_any_loop over the adaptive window with a pagination guard (state moved AND not a state-loop → don't fire).

GymRunner._is_loop is the single internal hook the runner consults for every soft-nudge / hard-terminate / loop-recovery / claude-director gate. It dispatches to is_any_loop_adaptive by default and falls back to the legacy is_any_loop when MANTIS_LOOP_ADAPTIVE=disabled.

Threshold logic

For the default soft=3 / hard=8 runner windows:

Pattern in last base samples	Effective window	Rationale
Diversity ≥ 0.6 (varied action types/coords)	base + 2	Productive exploration — defer the nudge.
State progressed (URL or frame moved)	base + 2	Pagination / drilldown — defer.
Diversity ≤ 0.25 AND state frozen	max(floor, base − 1)	Clear stuck signature — terminate sooner.
Buffer < base samples	base (unchanged)	Not enough history to judge.
Otherwise	base (unchanged)	No clear signal.

The pagination guard inside is_any_loop_adaptive provides the second half of the fix: even after extension, identical-click pagination would still trip is_repeat_loop. The guard short-circuits to False when state moved across the (effective) window AND is_state_loop would not fire.

Ablation toggle

Var	Default	Effect when set
MANTIS_LOOP_ADAPTIVE	enabled	disabled falls back to is_any_loop(base) everywhere — legacy fixed windows.

/v1/cua accepts loop_adaptive: true|false in the request payload to flip the toggle for that single request (the runtime patches os.environ under a try/finally since /v1/cua is serialized per container). Pairs with the existing per-request override surface established for perceptual_verify, loop_recovery, done_gate.

Use scripts/ablate_v1_cua.py --toggle loop_adaptive to run a paired A/B against a warm container.