Validate complete round sets and local round invariants

[komaksym]

Parent

• Implement auditable UFC dataset fidelity pipeline #14 — Implement auditable UFC dataset fidelity pipeline

What to build

Validate round observations as one complete set per stable fight rather than independently surviving rows. Establish structural and local round quality before comparing the set with fight totals.

Acceptance criteria

• A certified round set is evaluated at the complete-fight grain and requires stable fight_id; individual surviving rows cannot bypass set-level validation.
• Expected observed keys are exactly 1..finish_round when finish-round evidence is valid; missing tables produce unavailable, and missing expected rows produce incomplete.
• Conflicting (fight_id, round_number) observations, malformed/impossible values, and other local hard failures produce invalid and quarantine the entire round set while preserving every raw observation.
• Local arithmetic checks enforce integer landed/attempted pairs, 0 <= landed <= attempted, significant-strike target/position decompositions, integer control seconds, and percentage recomputation from counts.
• Schedule-derived checks run only when schedule metadata is interpretable; otherwise the relevant check is warning-level not_checked with a reason.
• A local round defect leaves an independently valid fight total usable as warning; it does not trigger cross-grain mismatch quarantine.
• Fixture tests cover unavailable, incomplete, invalid, duplicate-key, zero-activity, malformed, early-finish, three-round, and five-round sets.

Blocked by

• Capture canonical raw per-round UFCStats observations #16 — Capture canonical raw per-round UFCStats observations.
• Validate fight totals into trusted, warning, and quarantine states #17 — Validate fight totals into trusted, warning, and quarantine states.