feat: add turn-indicator false-positive (spurious-transition) rate metric

[MasatoSaeki]

Summary

Adds a closed-loop turn-indicator false-positive (spurious-transition) rate metric, the complement of the existing turn_indicator_transition_accuracy (#403).

• The existing metric answers: "when GT c?" (gated on a real GT transition,transition-based, not per-step).
• This metric answers the mirror question: "when GT holds steady, does the model flip anyway?"

Both are computed in the same _score_turnpopulation of real-inference (scored) steps, and are mutually exclusive per step: transition_total + fp_total always equals the number of scored steps a segment.

Metric details

• Scope: every scored step where GT did nscored step — regardless of which class(NONE/DISABLE/LEFT/RIGHT) it's holding, not restricted to None.
• Event: a "false positive" is the model'g from its own previous scored prediction(turn_indicator_prev_scored_pred), not onstruction in this branch, so comparingagainst GT would just remeasure transittead of measuring spurious model behavior.
• New _SegState fields: turn_indicator_fpl (accumulators) and turn_indicator_prev_scored_pred (baseline, lifecycle mirrors turn_indicator_prev_scored_gt: seeded in _seed_state, updated every scored step,port so the environment jump itself isnever counted as a spurious flip).
• turn_indicator segment-row block extendtransition_total} to also carry {fp_count,fp_total}; the false_positive_rate ratien fp_total == 0) is derived at aggregatetime, same convention as transition_acc
• Held/cached-plan steps (_hold_turn_indicator, no fresh inference) never touch the new counters, same as the existing ones.
• Wired through closed_loop_eval.aggregate/format_summary_lines, closed_loop_score_keys.SCORE_EXTRACTORS, scenario_sim_metrics (zero-count block,at path), wandb_closed_loop (new tablecolumn + weighted-by-fp_total aggregatid_loop._write_groups_manifest.

Test plan

• pytest scenario_generation/tests/test_turn_indicator_scoring.py -v — new tests cover: FP counted only whenGT is steady and the model's own predicsteady and prediction is stable;GT-transition steps never touch the FP pairs partition the scored steps,verified via transition_total + fp_totas don't touch the FP counters; unstickteleport re-seeds turn_indicator_prev_scored_pred.
• pytest scenario_generation/tests/te— turn_indicator_block (4-arg), aggregatepooling of fp_count/fp_total/false_posi → None guard.
• pytest scenario_generation/tests/te.py -v — fixture updated to the new blockshape.
• pytest scenario_generation/tests/ - pre-existing failure(test_wandb_scenario_sim.py::test_serieonfirmed unrelated to this change(reproduces identically on the unmodifing order-dependent mock-state issue).

🤖 Generated with Claude Code