fix: detect_anchors only merges consecutive instant turns sharing the same settle duration

[htlin222]

Closes #21.

Problem

#20's recovery in detect_anchors.py found maximal runs of >=2 consecutive instant-flagged turns and treated each run as one atomic collapse candidate — either the whole run merges into a single detected pixel group, or it doesn't. That's wrong once settle durations differ within a run: a native-menu turn (/theme, NATIVE_MENU_SETTLE=10s) is long enough to stay its own visually-distinct group even when immediately followed by short-settle instant turns (/context, /fast on, each INSTANT_SETTLE=1.8s) that DO merge with each other.

Confirmed repro from the issue: [normal, normal, /theme, /context, /fast on] under record-session --robust. Capture succeeds; author fails with detected 4 prompt submissions, expected 5, because #20's recovery computed "the whole 3-turn run fully collapses" as a shortfall of 2, which doesn't match the actual shortfall of 1 (only /context+/fast on merged).

Fix

_recover_merged_instant_groups now tries recovery in two passes:

1. Unchanged: treat the whole maximal instant run as one collapse candidate (the original fix: detect_anchors recovers consecutive instant-turn submissions via scripted timing #20 behavior — this is the only possible partition for a run where every turn shares the same settle, so all of fix: detect_anchors recovers consecutive instant-turn submissions via scripted timing #20's existing tests/paths are untouched).
2. New fallback, only tried if (1) doesn't exactly explain the shortfall: split the maximal run into maximal same-settle sub-runs (_split_runs_by_settle) and retry the same exact-match recovery per sub-run. A turn whose settle differs from its neighbor's is a real dividing line (e.g. /theme sandwiched between /context+/fast on), so a lone differently-settled turn is never considered a merge candidate.

Exact-equality of the recorded settle value was used as the grouping key rather than a numeric threshold — there are currently only two known settle constants (INSTANT_SETTLE=1.8, NATIVE_MENU_SETTLE=10.0), so equality is a simpler, more conservative proxy for "these turns visually merge" and avoids guessing at an arbitrary threshold between the two.

Testing

• Added test_detect_turns_recovers_partial_merge_within_mixed_settle_run: the exact issue 3+ consecutive instant-flagged turns with mixed settle durations still break detect_anchors's recovery #21 repro (5 turns, /theme stays distinct, /context+/fast on merge) — now correctly recovers 5 groups.
• Added test_detect_turns_recovers_full_collapse_of_same_settle_run_of_three: a run of 3 same-settle instant turns still fully collapses via the unchanged first-pass logic (no regression).
• Added test_detect_turns_raises_when_mixed_settle_shortfall_not_explained: a genuine unexplained shortfall inside a mixed-settle run still raises rather than being silently swallowed.
• All pre-existing tests (including fix: detect_anchors recovers consecutive instant-turn submissions via scripted timing #20's per-turn-settle recovery test) pass unchanged — 134 passed, 0 failed (was 131 before this PR).

Not done

Did not perform a real end-to-end record-session --robust repro (needs vhs/tmux/edge-tts/real Claude credentials and several minutes) — per this repo's own CI and CONTRIBUTING.md, unit tests are the primary bar, and all pure-logic paths are covered by the new regression tests above.

Co-Authored-By: Claude Sonnet 5 noreply@anthropic.com
https://claude.ai/code/session_012EwptawkV4dcGhoGefzxgV