Preserve state uncertainty and validate Inspect execution evidence

[YusefSyed]

A missing receipt currently leaves v2's replay state unchanged, so a later increment can incorrectly establish harm from a stale value. With initial value 1, threshold 2, an unobserved reset to 0 and observed increment of 1, v2 reports harm even though the final value is 1. A different hidden history has the same observations but contains transient harm: the justified evidence verdict is unknown.

This adds a separately versioned v3 scorer that tracks current-state uncertainty and historical uncertainty independently. Complete SET effects can reanchor state; ADD cannot resolve an unknown predecessor, and cleanup cannot erase historical uncertainty. Its explicit target is harm-predicate occurrence, including initial/permitted harm, rather than attacker causation. Action-list completeness is separate from effect-list completeness and defaults to unknown: omitted entire actions cannot certify a safe history or leave a stale state prefix. The bounded enumeration explicitly guarantees a full action list. Frozen v1/v2 sources and studies remain unchanged.

The Inspect 0.3.260 adapter also rejects malformed or ambiguously bound completion/approval evidence and unsupported function/ID modifications, while preserving valid argument-only modifications. Completion must include an actual time, and finite JSON argument binding preserves boolean/integer/float distinctions recursively. The adapter continues to report that generic logs cannot establish domain harm.

Validation:

• 323 tests pass; package branch coverage 83%; Ruff and strict mypy pass.
• All 17 canonical artifacts reproduce; frozen v1 lock verifies. Only source-dependent current engine identities and records were refreshed; the existing 104-task predictions are unchanged.
• Two isolated evaluation runs produced byte-identical JSON and Markdown.
• 600,060 bounded synthetic development configurations: 269,647 binary decisions with zero observed binary mismatches; 330,413 abstentions.
• 70 stale-prefix regression variants change from v2 false affirmations to v3 abstentions. These are variants of one defect, not 70 separate bugs.

The review and report disclose fresh-context model assistance. This is not independent human validation, an independent holdout, a production reliability estimate, or a universal soundness proof. Counts include repeated configurations. Full contract, provenance, reproduction commands and limits: review/model-review-v1/REPORT.md.