F450 MULTI-ROUND REVIEW RECORD
Captured: 2026-08-25 UTC

R1 - Trust boundary and transitive closure
Reviewed root-record normalization, proof identity, recursive imports, store
integrity and the cache-hit linearization point. Rejected the first design
because a cached parent could have hidden later mutation of an imported CAS
object. The final entry carries every transitive dependency's encoded-byte
SHA-256 and re-attests the entire closure before authorization.

R2 - Protocol identity and persistent compatibility
Reviewed proof policy, receipt identity, store identity and configuration
drift. Rejected a version that folded cache budgets into the existing checker
policy because it would invalidate persistent artifacts for a local
optimization. The external ordered-LRUP policy remains stable; F450 has a
separate cache protocol and policy digest.

R3 - Performance counterexample and resource bounds
Measured the initial recursive-JSON implementation and the per-object
attestation revision; both had approximately 1x replay ratios because each
dependency reopened flock/SQLite and reparsed JSON. Replaced them with one-lock,
one-connection batch attestation. Entry-count and accounted-byte LRU budgets,
oversized rejection and hard configuration maxima bound resident state.

R4 - Mutability, object lifetime and concurrency
Reviewed Python object-id reuse, plan mutation, weak references and concurrent
first-use races. Cache admission now requires an exact BitBlastPlan with tuple
clauses, integer literals and frozen CnfIncrement records. Weakref plus replay
scope prevents stale id reuse. A 16-thread adversarial test requires one stable
authorization identity and a single resident entry.

R5 - Failure semantics, telemetry, tests and claim boundary
Found that the initial hit counter advanced before dependency attestation, so a
tampered closure was counted as a successful hit. The final implementation
increments hits only after complete attestation, records attestation failures
and object/byte work, and evicts exactly the failed entry. Documentation and
the oracle distinguish proof-replay mechanism ratios from solver, coverage and
defect-yield claims.

FINAL GATES
- Focused: 9 passed.
- Coupled regression: 88 passed, 50 subtests passed.
- Canonical inventory: 1428 node IDs; exactly 9 F450 additions, zero removals.
- Full capability gate: 1428 passed, 310 subtests passed; zero skip, xfail,
  xpass, deselect, collection error, missing nodeid or unexpected nodeid.
- External capabilities: 16/16 present.
- Static checks: recorded in static-checks.txt.