[Bug]: adaptive replay planner accepts NaN replay-cost values #6094

Preflight

• I searched existing issues and discussions for this behavior.
• This report is not an unpatched security vulnerability.
• The report and attachments contain no credentials or private/internal material.

Issue origin

Inferred by an AI agent and not yet observed or reproduced

LoopX version or commit

bd51193

Host or runtime surface

Codex App

LoopX area

Control plane (goals, todos, quota, scheduler, registry, runtime)

Problem

ReplayPositionCost validates replay costs with _nonnegative, which only rejects values where normalized < 0. Since NaN < 0 is false, a NaN capture or restore cost passes validation and reaches proposal scoring; the resulting NaN score is then admitted because NaN < minimum_score_minutes is also false. The proposal list can therefore contain non-finite scores and sort unpredictably. The validator also accepts booleans and leaks OverflowError for integers too large to convert to float.

Minimal reproduction

Construct a ReplayPositionCost with capture_minutes=float("nan") and restore_minutes=0.0, then pass it with a valid FailureReplayObservation to plan_adaptive_replay_points. The cost is accepted and the resulting proposal can have score_minutes=nan. The current helper expression also returns nan for this input.

Expected behavior

Replay costs must be finite non-negative numbers. Reject NaN, infinities, booleans, and values that cannot be converted to float with a field-specific ValueError.

Actual behavior

The cost validator accepts NaN because its negative-only check does not reject non-finite values. Scoring then admits the NaN result because comparisons against the threshold are false.

Sanitized diagnostics

Additional context

No response