Evidence conformance and composed guarantees

C8 adds test/example infrastructure, not a production backend or a new inspector. The reusable runner-independent harness is tests/conformance/evidence.ts. runEvidenceConformance(profile) accepts a scoped adapter factory and declared acknowledgement strength, with optional same-backend reopening. It uses Node assertions and no test-runner imports. It is not exported or eagerly imported by the production package. Existing adapter-specific tests remain.

Supported tested profiles are the existing process-local InMemoryEvidenceStorage, FileEvidenceStorage restart acknowledgement and file fsync acknowledgement. The harness checks immutable admission/read copies, exact bounded UTF-16 reconstruction, source hashes, corrupt references, tenant isolation, expiration/deletion/pruning, quota admission and concurrent immutable writes. Reopening an in-memory store is not restart support. Local fsync does not establish replicated durability. No shared Mongo/SQL backend profile has been delivered or claimed.

tests/fixtures/composition-evaluation/restart.mjs runs as two actual Node processes. The producer stores remote source and proxy observations, then exits. Recovery uses the same scoped file backend, restores exact quoted data and proxy failure, denies unauthorized access, and observes a new server epoch as unknown. It preserves one synthetic submission and refuses replay. Evidence recovery does not restore a routine continuation, live ModelTask, authority or workflow cursor.

Run npm run example:composition-evaluation for six local/remote meeting workflows with development and held-out data. Each reads exact source, submits once, waits for host confirmation and produces a temporary artifact checked against the source. The model's acknowledgement and accepted submission are insufficient for completion. Rejected continuation authorization performs no work. Artifacts are deleted during cleanup. The example reuses the existing canonical baseline reporter utility and ordinary public APIs. Reported latency is a synthetic functional observation, not a comparative quality/performance claim. Known model cost subtotal, unknown model pricing and unknown external costs remain separate; missing cost is never zero.

Existing C3–C7 tests remain the authority for uncertain writes, cancellation and late children, long-source projection/compaction, complete pagination and connector scope. C8 runs those suites alongside the new composed and child-process cases. The stable collection fixture includes empty pages, unread tails and a final partial page. Do not infer corpus coverage from returned count or provider pagination end.

For inspection, compose agent.inspectCapabilities(), runtime inspection, C3 runner.inspect(operationId), C6 child observers, C7 session.inspect() and explicit ModelTask snapshots. Old report shapes are unchanged. Configured host checks are separate from observed host verification. Managed, direct, Realtime and external loops keep their existing owners. Remote observations do not prove business completion; compaction/cancellation support is not proof an external write stopped. Unknown/custom support stays unknown. There is no universal execution registry.

AMOS /routine inspect displays observed verification, accounting binding, recorded call count or unknown, and same-process continuation limits. Run npm run example:amos-composition for its offline actual-command demonstration. No new voice/remote transport UI is introduced. ICOS adoption, releases, new database adapters and durable workflow resume remain separate work. Live LLM testing is paused at the user's request as of 2026-10-28.

Runtime harness adapters must retain the original C7 ToolManager and evidence handle and preserve raw tool returns independently of normalized outcomes. The existing conformance profiles remain the evidence contract; runtime control records are separate.

H3 recovery checks include actual producer/recovery processes, original-result redelivery, write-before-crash receipt reconciliation, foreign-scope denial and acknowledgement-failure retention. They exercise control durability separately from the existing evidence conformance profiles. Neither mock service receipts nor runtime metadata assert business verification or restore workflow authority.