Contract-Parity Audit

Two matrices answer one question: where does a shared contract get advertised but not delivered? Matrix A diffs every method of a multi-impl interface across its implementations. Matrix B maps every public surface exposing a given operation to the chokepoint it actually reaches. Do not redesign anything — the deliverable is the enumerated gap set that scopes the fix.

Provenance: distilled from the rosary-46e7ff run (2026-09-02, archive ~/codebase-audits/agentic-research-rosary/3274bd9/contract-parity/) that scoped rosary-44eec8/45504f/457927: the parent audit had sampled 4 methods and found 2 divergences; the full matrix found 9 — a hit-rate lesson in why sampling doesn't scope a fix. Sibling of counterfactual-audit; shares its iron rules (durable evidence directory FIRST, OBSERVED/INFERRED tagging, read-only, fresh-context subagents, thesis withheld).

Setup

Durable output before any dispatch: ~/codebase-audits/<org>-<repo>/<commit>/contract-parity/ — one file per matrix. Resolve the full commit SHA; note whether a parent audit exists and link it. Findings are banked as issues after, never during.

Matrix A — implementation parity

For EVERY method of the target interface (enumerate from the definition — no sampling), one row, reading each implementation end-to-end including delegation chains:

1. Method + production callers ("trait-dispatch only" is an answer).
2. Per-impl location (file:line) + one-line behavior.
3. Validation performed, per impl.
4. Atomicity/concurrency mechanism, per impl (lock-held sequence / single statement / transaction / non-atomic multi-step).
5. Sanitization (secret scrubbing etc.), per impl.
6. Event/audit side effects, per impl.
7. Error posture (propagated vs logged-and-continued vs silent), per impl.
8. Idempotency / retry safety, per impl.
9. Deliberate-bypass semantics (escape hatches are design, not gaps — verify against the incident they exist for before "fixing" them).
10. Tests per impl (name them, or NONE).
11. The punchline: observable semantic difference.

Classify each method as exactly one of:

UNTESTED may co-occur with any class; list those methods separately. Never assume identical implementations are required — the hunt is for silently weaker semantics, not for textual sameness.

Matrix B — surface chokepoints

Enumerate every externally reachable surface that performs the operations under audit (CLI verbs, RPC/MCP tools, HTTP routes, webhooks, IPC, worker loops, internal sweeps). Per surface-operation row: entrypoint (file:line); parsing/validation; the application function reached (or "direct store call"); lifecycle/policy gates applied; verification performed; persistence ops; retry/idempotency; result reported to the caller and whether failure can be mis-reported; tests.

Then group by OPERATION and state, per group, whether all surfaces share one chokepoint — naming every bypass with file:line, and whether each bypass is sanctioned (its own gate + suppression annotation + explanatory comment) or a defect.

Calibrations (each earned in the reference run)

• Verify consumer liveness before classifying severity. The reference run's parent audit called a lossy mapping "live on every sync pass"; tracing its consumer chain found zero production callers — the live path used a different, correct mapping. Wrong liveness = wrong priority.
• Enumerate, don't sample. 2 divergences in 4 sampled methods became 9 in 34. A sampled matrix cannot scope a fix.
• Spot-verify before building on a row. The lead re-read source for 6 matrix claims before any implementation started; all held — but the skeptic's later "no open issue references X" claim failed because it searched the repo tree and not the issue tracker. Search both.
• Grep discipline: unqualified/local call forms (use super::* callers of a bare fn), and both directions (who calls this / what does this call).
• Fixture blind spots are findings. The reference run's handler tests silently exercised an empty pool for years (a test constructor seeded one map but not the one the handler read) — "tests exist" rows deserve one check that the test can fail.
• Fix framing for the gap set: same policy, per-impl-appropriate mechanism — copying one impl's lock pattern onto another's connection pool is a race that claims the guarantee without delivering it. Never prescribe a shared wrapper across impls with different concurrency models.

Deliverable and follow-through

End each matrix file with: (a) the CONTRACT VIOLATION gap set as a numbered list; (b) the UNTESTED list; (c) limitations. Bank the gap set as a comment on the fix-owning issue — acceptance criteria must carry the mechanism constraints and every deliberate-bypass carve-out (with its incident id), so the fix cannot "close the gap" by breaking the escape hatch. Route findings that belong to other issues explicitly; file new issues for unowned ones after the audit, dedup-checked.

Common mistakes