chore: refuse false benchmark completion

[abrichr]

Summary

• reject malformed actions instead of treating them as task completion
• separate agent failures from unavailable infrastructure and evaluator attempts
• require explicit action, setup, and evaluator receipts before a result can count
• exclude unavailable attempts from measured outcome rates while preserving their diagnostics
• reject malformed persisted results and success-shaped report values

Why

The evaluation stack could turn parser failures, transport failures, missing evaluator responses, and malformed result artifacts into plausible completed outcomes or measured zero scores. This change keeps those states explicit and prevents them from entering success metrics.

Validation

• 2,136 light tests passed, with 55 skipped and 17 deselected
• 100 focused evaluator and receipt tests passed after the final receipt correction
• Ruff passed
• the wheel and source archive passed the public source-policy boundary check
• git diff --check passed

Release

This merge does not publish a package release.