F452 MULTI-ROUND REVIEW RECORD
Captured: 2026-08-26 UTC

R1 - Upstream contract and trust boundary
Pinned ImpCheck incremental commit b5f37b21385ee802ce015103b23aff62f92b1734
and reviewed src/trusted/clause.c. Preserved signed-to-unsigned ordering,
sorting, delta coding, canonical 7-bit variable integers, total-size prefix,
and the seven-byte inline threshold. Compression remains an internal
representation after proof authorization, not a new proof rule.

R2 - Integer, encoding, and storage safety
Rejected zero, INT_MIN, duplicate literals, tautologies, noncanonical varints,
truncation, size mismatch, uint32 accumulation overflow, and resource excess.
Used explicit array/owned heap storage rather than pointer representation.
Fixed the empty C ABI decode path so it never passes a null output iterator to
the generic copy algorithm. Insufficient output capacity performs no write.

R3 - Production hot path and ownership
Replaced import-queue and activity-tracker integer vectors with move-only
CompressedClause objects. CaDiCaL callbacks decode one literal at a time.
Semantic queue capacity remains based on literal count; encoded-byte accounting
is independent. A decode failure requests termination, rejects the import, and
cannot produce an ACK.

R4 - Configuration, ABI, and backward compatibility
Added a versioned protocol and complete C ABI checks. Found that the new
require_clause_compression boolean was mistakenly included in numeric option
validation; separated strict boolean validation and added a regression. Old
shims remain allowed by default, while explicit require fails closed.

R5 - Telemetry and persistence
Found native compression counters were visible only in the live context and
would be discarded from stored results. Added monotonic session deltas and
QueryStore validation for complete protocol-bound telemetry, inline/heap
conservation, queued bytes, decoded literals, and zero compression failures.

R6 - Tests, sanitizers, performance counterexamples, and claims
Added exact bytes, extremes, corruption, capacity, property, concurrency,
configuration, ABI, oracle-tamper, and store-tamper tests. A test insertion
initially moved two cleanup lines into the next unittest; restored the original
lifetime and reran coupled/full gates. ASan/UBSan completed 100000 random native
round-trips. The report retains the 11.4% one-literal payload expansion and
does not relabel payload reduction as RSS, solver, coverage, or defect-yield
improvement.

FINAL GATES
- Focused: 23 passed.
- Coupled compression/realtime regression: 37 passed.
- Canonical inventory: 1467 node IDs; exactly 24 additions, zero removals.
- Full capability gate: 1467 passed, 310 subtests passed; zero skip, xfail,
  xpass, deselect, collection error, missing nodeid, or unexpected nodeid.
- External capabilities: 16/16 present.
- Static and sanitizer checks: recorded in static-checks.txt.