Make BTC 5m settlement probability a market-level research target

[proerror77]

Parent

#314

Change contract

Make final settlement prediction an explicit market-level task keyed by market_id and decision time, with p_up as its prediction and official Polymarket resolution as its label. Remove execution PnL from the settlement task identity without changing token execution calculations.

Acceptance evidence

• The typed mission or evaluator task identifies settlement probability independently from token execution.
• One sample is keyed by market_id and decision time; Up and Down are mutually exclusive sides of the same settlement.
• Metrics are event-disjoint Brier score, log loss, calibration, and settlement verdict evidence.
• Up/Down fill, slippage, and repricing fields cannot become settlement labels.
• Common-time cohorts keep overlapping events on one side of every split.
• Existing full-surface and baseline source gates remain fail closed.
• Focused counterexamples reject token labels, row-random splits, and cross-cohort leakage.

Out of scope

Token execution task design; new horizons; collector or snapshot changes; MCTS controller migration; Paper or Live; profitability claims.

Dependency and merge order

Depends on #321 and #322 so the settlement task is expressed through Mission v3 and the shared immutable event partition. It may merge before the token execution task and must not absorb baseline eligibility or collector work.

Focused validation

Use TDD at the mission, task identity, split, and settlement evaluator boundaries.

Rollout and rollback impact

Research-only schema or task-contract change. Rollback restores the prior mixed target identifier; immutable old artifacts remain versioned and are never silently reinterpreted.

[proerror77]

Completed by the merged settlement-isolation implementation beginning at 33b1d58c8e899aa7be9001d342b33909df45ff9e (ancestor of current main): Mission v3 has a distinct settlement task; settlement verdicts use exactly one Up/event-root sample per decision and event-disjoint common-time splits; Brier/log-loss/calibration are computed from the official settlement label; execution evidence cannot change settlement feedback. Revalidated on current main with factors_v2::tests::settlement_verdict* (2 passed) and prediction_feedback_is_byte_identical_when_execution_evidence_changes (1 passed).