[Aparnap2]

What changed

Two files, test-only:

File	Change
fixtures/golden_cases.json	+10 cases appended (GOLDEN-10..19). Cases 01–09 byte-identical (machine-verified), version unchanged at 1.0.0, nothing removed, no existing expected value rewritten.
apps/api/internal/claims/parity_test.go	+750/−15: a coverage registry replacing the hard-coded count + prefix list, a two-way fixture↔registry contract test, 9 new helpers, 10 new tests.

No production Go file, Python artifact, CI workflow, dependency, or doc changed. Base: main@254c4b92dddf59a758a133c1edbf7618dcc011b2. Commit: df3309a.

The gaps this closes

MALFORMED_CLAIM, TENANT_MISMATCH, INVALID_EVENT_ID and UNKNOWN_STATUS were referenced by zero Go test files. Six legal edges had no asserted coverage either — because ClaimStatusInvestigationRequired and ClaimStatusWaitingForEvidence appeared in no test file at all.

Re-derived edge counts (the previously reported "8 of 13" was wrong on both sides): the graph has 14 legal edges; 4 were asserted in any Go file before this change (one of them PostgreSQL-gated only); 10 were asserted nowhere; one of those is covered by open PR #178, leaving 9 genuinely uncovered on this base. This change covers 6 of the 9. DB-independent legal-edge assertions go from 3 → 9.

Case	Covers	Expectation
GOLDEN-10-malformed-claim	MALFORMED_CLAIM (was 0 coverage)	empty status / version < 1 / empty tenant → MALFORMED_CLAIM; malformed precedes tenant mismatch
GOLDEN-11-tenant-mismatch	TENANT_MISMATCH (was 0)	acting tenant ≠ claim tenant → TENANT_MISMATCH, and it wins over blank event-id, stale version and illegal edge
GOLDEN-12-invalid-event-id	INVALID_EVENT_ID (was 0)	whitespace-only and empty idempotency key → INVALID_EVENT_ID; wins over version and legality
GOLDEN-13-unknown-status	UNKNOWN_STATUS (was 0)	unrecognized source or target → UNKNOWN_STATUS, not ILLEGAL_TRANSITION; version check precedes the status check
GOLDEN-14-validating-to-exception	VALIDATING → EXCEPTION	success, EXCEPTION, v1→v2, key recorded, input unmutated, one audit event
GOLDEN-15-exception-to-investigation	EXCEPTION → INVESTIGATION_REQUIRED	INVESTIGATION_REQUIRED, v2, one audit event
GOLDEN-16-investigation-to-waiting-evidence	INVESTIGATION_REQUIRED → WAITING_FOR_EVIDENCE	WAITING_FOR_EVIDENCE, v2, one audit event
GOLDEN-17-waiting-evidence-to-hitl	WAITING_FOR_EVIDENCE → HITL	HITL, v2, one audit event
GOLDEN-18-hitl-to-action-pending	HITL → ACTION_PENDING	ACTION_PENDING, v2, one audit event
GOLDEN-19-action-pending-to-actioned	ACTION_PENDING → ACTIONED	ACTIONED, v2, one audit event

Every expectation is derived from the documented check order in statemachine.go (malformed → tenant → event-id → idempotent replay → version → known status → legality), the transition graph, and the canonical chain in docs/AGENT_OPERATING_CONTRACT.md. No business rule was invented.

The harness no longer maintains a second count

goldenExpectationByID is a 19-entry coverage registry of hardcoded literals, matched against the fixture in both directions: every fixture id appears exactly once, every registered id is present exactly once, every fixture case has Go coverage, and every case's fixture expected equals the hardcoded Go expectation. A fixture case without coverage now fails (fixture case %q has no Go coverage) instead of passing silently. The previous len(doc.Cases) != 9 assertion and the nine-entry prefix list are gone.

Each new case re-asserts the fixture's inputs against hardcoded literals, runs the scenario through the direct Go APIs, asserts the observed outcome against hardcoded literals, then asserts the fixture's expected against the same literals. Proven live rather than echoed: mutating the fixture broke the tests, and separately mutating the Go expectation broke them with the observed value reported — the expected side is real implementation output, not a fixture echo.

Verification

Run by me against real PostgreSQL (claimops-postgres, claimops_app with rolsuper=f, rolbypassrls=f, so RLS is genuinely enforced):

go test ./... -count=1                        -> 41 packages ok, 0 failures   (4 more packages ran than without a DSN)
go test -race -count=1 ./...                  -> 37 packages ok, 0 failures, no data races
go vet ./... / go build ./...                 -> clean
go test ./internal/claims/... -v -count=1     -> 24/24 PASS, 0 skips

The PG-gated suites that the authoring agent could not reach were run for real and pass: TestClaimSubmit_* (6/6), TestDecision_ExpireFromHITL_Live, TestDecision_ExpireFromNonHITL_Rejected_Live, TestDecision_DuplicateIsReplay, TestDecision_ConcurrentIdenticalOneMutation, TestDecision_ConcurrentDifferentEventsConflict, TestHITL_ExpiredExcludedFromPending.

Not run (services absent, not silently skipped): Mockoon contract tests (gate returns 404 on GET /v1/policies/POL-001), localgcp GCS/Pub/Sub, GCW emulator, and the Groq live-LLM matrix. The authoring agent's gofmt binary and go clean -testcache were blocked by its shell policy; formatting was verified with go fmt plus a read-only go/format equivalent, and every run used -count=1, which bypasses the test cache. uv/ruff/pytest were not run — no Python changed.

Findings — deferred, not encoded

• F1 (doc defect, out of scope): statemachine.go:12-14 claims golden case 08 witnesses "stale-version wins over legality", but case 08's edge REGISTERED → DOCUMENTS_RECEIVED is legal, so it witnesses only that a stale version is rejected. The precedence is real (check 5 precedes check 7) but unwitnessed in Go.
• F2 (new cross-implementation divergence): the fixture's expected.audit_actions is ["claim.status.transition"] — the Python action string — while Go emits "claim.transitioned". Asserting the fixture's value in Go would fail. The new cases therefore use audit_count (language-independent: both implementations record exactly one audit event per successful transition) instead of the action string.
• F3 (terminal_state deferred): the field is read by neither implementation and its meaning differs across cases; GOLDEN-01's value ("VALIDATING") contradicts the outcome both suites assert (REGISTERED, v2). Correcting it would require ratifying a definition no ADR or contract establishes, so it is documented in the parity_test.go package header rather than silently rewritten. (The one contract-supported correction would be GOLDEN-01 → "REGISTERED"; say the word and it can be applied with the reasoning stated.)
• F4 (new Go/Python definition divergence): Go treats a malformed claim as Status == "" || Version < 1 || Tenant == ""; the Python spec tests absence (None), not emptiness, and its model makes version < 1 and a blank tenant unconstructible, so status="" would be UNKNOWN_STATUS there rather than MALFORMED_CLAIM. GOLDEN-10 therefore encodes the authoritative Go definition and cannot be satisfied by the Python spec artifact. This warrants its own issue.
• F5: no Python test asserts GOLDEN-10..19, so those ten are currently a Go-side contract. Deliberate and stated — this change is Go-scoped.

Residual, not silently dropped: edges REGISTERED → DOCUMENTS_RECEIVED, DOCUMENTS_RECEIVED → VALIDATING and VERIFIED → CLOSED remain unasserted; ACTIONED → VERIFIED remains PostgreSQL-gated; the STALE_VERSION-before-ILLEGAL_TRANSITION precedence remains unwitnessed (F1); READY_FOR_REVIEW terminality remains delegated to PR #178 (no overlap with this change — no terminality assertion and no READY_FOR_REVIEW edge here, and no symbol collisions).

Status

Test-only, deterministic (fixed literal ids, fixed literal dates, exact int64 paise; no uuid4(), no time.Now(), no randomness, no environment-dependent expectation, no secrets, no patient data). Not authorized to merge — opened as a draft. Release status NOT QUALIFIED; this change qualifies nothing and closes no entry in docs/KNOWN_LIMITATIONS.md.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: b3afd38c-72da-46ea-853a-f49ed2515fae

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}