P0.5.2 — δ Fallback Chain — Verification

Step 5 of the 5-step chain. Test evidence for the Phase 0 library-only single-member fallback stub mandated by ADR-005 §Decision. Closes Phase 0 at 28/28.

1. Gate results

cd .worktrees/claude/p0-5-2-fallback
npm run build       # clean (no TS errors)
npm run lint        # clean (no ESLint warnings/errors)
npm test            # 26 suites / 1084 tests passing

Test count comparison

Metric	Before P0.5.2	After P0.5.2	Δ
Test suites	25	26	+1 (src/__tests__/domains/router/fallback.test.ts)
Tests	1051	1084	+33

The +33 (vs. packet target of 10–15) exceeds the lower bound because two additional describes were added during impl: routeRequest — upstream forwarding subtests (seam-injection passes) and default dispatcher tests (coverage 100% on fallback.ts). All 33 tests pass first try after the initial ESLint curly fix.

Coverage (fallback.ts, focused)

-------------|---------|----------|---------|---------|-------------------
File         | % Stmts | % Branch | % Funcs | % Lines | Uncovered Line #s
-------------|---------|----------|---------|---------|-------------------
 fallback.ts |     100 |    85.71 |     100 |     100 | 217-219, 310
-------------|---------|----------|---------|---------|-------------------

• Statements: 100%
• Branches: 85.71% — uncovered are optional-chain / nullish-coalescing defenses in the FallbackChainExhaustedError ctor’s attempts[N-1] access (N is always ≥ 1 in Phase 0 so the branch .error.message ?? 'unknown' is defensively untestable).
• Functions: 100%
• Lines: 100%

2. Acceptance criteria verification

Every AC from the contract §6 is verified by at least one passing test.

AC	Contract ref	Test case in fallback.test.ts	Status
AC1	happy path returns RouteResult	returns RouteResult with model='claude' and content on success + projects upstream CompletionResult fields into RouteResult + RouteResult.model is abstract 'claude', not upstream version string + RouteResult is frozen	PASS
AC2	delegates to scoring	consults scoreIntent before calling upstream + uses default scoreIntent when scoringFn not injected	PASS
AC3	delegates to upstream (1×)	calls completionFn exactly once on success	PASS
AC4	forwards prompt	forwards prompt unchanged to upstream completionFn	PASS
AC5–AC6	forwards maxTokens/systemPrompt/model/apiKey	forwards maxTokens / systemPrompt / model / apiKey to upstream + omits optional keys from upstream options when not provided + forwards injected fetchFn / logger / delayFn to upstream	PASS
AC7	FallbackChainExhaustedError on failure	throws FallbackChainExhaustedError when upstream throws AnthropicApiError	PASS
AC8	attempts.length === 1	FallbackChainExhaustedError has exactly 1 attempt	PASS
AC9	attempts[0].model === ‘claude’	attempts[0].model === 'claude' (Phase 0 invariant)	PASS
AC10	preserves AnthropicApiError	preserves AnthropicApiError in attempts[0].error (same instance)	PASS
AC11	preserves AnthropicConfigError	preserves AnthropicConfigError in attempts[0].error (same instance) + exposes cause pointing to the last attempt error	PASS
AC12	ZERO cascade	does not retry after upstream failure — completionFn called exactly 1× + does not retry after AnthropicConfigError	PASS
AC13	determinism	two identical calls route to the same model + different prompts route to the same model (Phase 0 single-member)	PASS
AC14	ROUTER_PHASE_0_SHAPE marker	asserts single-member chain (members === 1) + asserts no circuit breaker (hasCircuitBreaker === false) + asserts modelsSupported contains only 'claude' + is deeply frozen	PASS
AC15	tools passthrough	passes tools array through to completionFn via default dispatcher + accepts empty tools array without error + dispatches to createCompletionWithTools when tools non-empty	PASS
AC16	error message format	includes attempt count and model name + error name is 'FallbackChainExhaustedError' + error code is 'FALLBACK_CHAIN_EXHAUSTED'	PASS
AC17	non-Error thrown values	wraps non-Error thrown string into Error instance	PASS

3. Invariant verification (ADR-005 §Decision)

#	Invariant	Evidence
I1	Exactly one upstream call per routeRequest, success or failure	calls completionFn exactly once on success + does not retry after upstream failure
I2	Winner always ‘claude’ in Phase 0	consults scoreIntent + uses default scoreIntent — both return ‘claude’
I3	RouteResult.model === ‘claude’ (abstract ID, not version)	RouteResult.model is abstract 'claude', not upstream version string
I4	Throws FallbackChainExhaustedError on failure	throws FallbackChainExhaustedError when upstream throws AnthropicApiError
I5	err.attempts.length === 1	FallbackChainExhaustedError has exactly 1 attempt
I6	err.attempts[0].model === ‘claude’	attempts[0].model === 'claude' (Phase 0 invariant)
I7	Preserves original error instance	preserves AnthropicApiError ... (same instance) + preserves AnthropicConfigError ... (same instance)
I8	No timers / circuit breaker state	Grepped fallback.ts — no setTimeout, setInterval, Date.now(), Date().
I9	No MCP tool registered	Grepped src/server.ts on worktree — no new imports from src/domains/router/*; tool list unchanged at 14.
I10	No env var reads	Grepped fallback.ts — no process.env.
I11	ROUTER_PHASE_0_SHAPE asserts invariants	All 4 shape tests pass.
I12	Deterministic routing	two identical calls route to the same model + different prompts route to the same model
I13	Tools passthrough dispatches to createCompletionWithTools	dispatches to createCompletionWithTools when tools non-empty

4. Forbidden-checks (§forbiddens from the task prompt)

Forbidden	Check	Status
No MCP tool registered	src/server.ts diff against main — no change	OK
No env vars	src/config.ts diff against main — no change	OK
No circuit breaker / timers / retries in Phase 0	grep -nE 'setTimeout\|setInterval\|Date\.now\|new Date' src/domains/router/fallback.ts → 0 hits	OK
No multi-model fallback	attempts.length === 1 test + ROUTER_PHASE_0_SHAPE.members === 1 test	OK
No real Anthropic API in tests	All tests inject completionFn OR a stubbed fetchFn	OK
No modification of scoring.ts	git diff origin/main -- src/domains/router/scoring.ts → empty	OK
No modification of src/domains/integrations/claude.ts	git diff origin/main -- src/domains/integrations/claude.ts → empty	OK
No modification of src/server.ts / src/config.ts	git diff origin/main -- src/server.ts src/config.ts → empty	OK

5. Non-regression

• All 25 pre-existing test suites remain green. Pre-existing intermittent startup — subprocess smoke flake (documented in memory as “flagged by Waves F and G sub-agents; predates”) intermittently surfaced once on first run; second + third runs clean. Not caused by P0.5.2.
• src/__tests__/domains/router/scoring.test.ts — 23 tests, all green — unaffected.
• src/__tests__/domains/integrations/claude.test.ts — claude wrapper tests unchanged.
• ESLint config + tsconfig unchanged.

6. Deviation from packet

The packet sketched 14 test cases; the actual file landed 33 because:

1. Multiple ACs collapsed into shared describes got split out for readability during impl (e.g. forwards X to upstream became three separate assertions + a negative test for omission + a separate seams test).
2. default dispatcher tests were added to lift fallback.ts coverage from 86.36% to 100% on stmts/funcs/lines — needed to exercise the createCompletion / createCompletionWithTools dispatch branch when completionFn is NOT injected.

Neither deviation changes scope; both strengthen evidence.

7. Phase 0 completion

This task is the last of 28 Phase 0 sub-tasks.

Phase 0 status	Before	After
Non-deferred sub-tasks complete	26/28	28/28
δ concept colibri_code	none (deferred)	partial (Wave I close commit will graduate)
Phase 0 shipped code axes	7/15 (α β γ ε ζ η ν)	8/15 (+ δ)
MCP tool surface	14	14 (zero added)

Per ADR-006 graduation rules, shipping a Phase 0 stub lifts the concept’s colibri_code field from none to partial. The frontmatter update belongs to the Wave I close commit, not this PR.

8. Conclusion

All acceptance criteria pass. All ADR-005 §Decision invariants are enforced by tests. Zero tool-surface, env-var, DB, or timer additions. Phase 0 closes at 28/28 on the boundary mandated by ADR-005. Ready to open PR.