test(acp): lock terminal correlation contract

[sethkarten]

Stack

Conformance child of #2320 only.

• V3 / base: test(acp): consume canonical correlation fixture #2320 (v080/acp-v3-correlation-final) at 799dd10e65921241773c2072e246dd07383a716b
• Conformance: 9204448b9746ac3b7deb9f72a4a618c6771009b4

Both PRs remain draft/HOLD. The upstream live-test environment is dependency-blocked; no live or paid evaluation was run.

Scope

One test-only conformance file. It preserves V3 observable metadata histories while requiring a single consistent producer evidence ABI for trusted terminal correlation. Stop reasons, arrival order, grace timing, opaque/partial claims, and pre-cut cancellation remain untrusted.

Verification

• independent read-only review: APPROVE
• exact one-path cumulative scope
• clean/diff-check
• consistent promptTurnId/eventSequence/phase/outcome ABI

Note

Add ACP 0.11 correlation conformance test suite for terminal pair scoring

Adds test_acp_correlation_contract.py to lock the observable-vs-trusted metadata contract and terminal pair scoring semantics for the ACP v1 adapter.

• Tests cover tool-only turns, late/opaque metadata arrival, error propagation, and producer evidence ABI validation.
• A load_runner_without_acp_dependency fixture monkeypatches minimal ACP doubles so the runner can be exercised in isolation.
• has_verified_producer_correlation and scoreable_terminal helpers encode the gate logic for trusted producer evidence and irrevocable completion-cut semantics.
• Cancellation durability tests assert that a sealed completion pair survives post-cut cancellation but is rejected if cancelled before sealing.

^{Macroscope summarized 9204448. (Automatic summaries will resume when PR exits draft mode or review begins).}