Score outpost-001 against the customer's endpoint, not one the agent picked

[leggetter]

Refs #47 — the first of the "does this scenario actually discriminate" items, and the one the code already knew about.

What was being measured

The ticket said "so they receive order events at their endpoint" and named no endpoint, so any reachable URL satisfied it. Across the six stored cells of this scenario the agents picked six different receivers:

arm	receiver	result
claude-code-sonnet-5	hkdk.events source it created	pass
claude-code-sonnet-5-no-skills	webhook.site inbox	pass
codex-gpt-5.6	hkdk.events source	pass
codex-gpt-5.6-no-skills	hkdk.events source	pass
codex-gpt-5.4-mini-no-skills	mock.hookdeck.com path	pass
codex-gpt-5.4-mini	localtunnel inside its own sandbox	fail

The failing cell is the one worth reading. It built the tenant, the destination and the topic set, published an order event, and reported delivery confirmed with 200 attempts. Its tunnel died with the container before scoring. So the check named "an order event reaches the customer" was discriminating on whether an improvised receiver outlived the run.

EVAL.ts already carried this as a known hole and deferred it, because closing it changes a published scenario.

Why now

Three things make this the cheap moment rather than the expensive one:

• publishing is held under Publishing is held: clear EVALS_PUBLISH when a clean snapshot lands #66, so no snapshot moves;
• the scenario fingerprint from Record what each published row was measured under #65 records the change for anyone comparing rows across it;
• the milestone-2 matrix has not been paid for. Landing this afterwards would mean measuring this scenario twice.

The change

The ticket names https://mock.hookdeck.com/api/v1/acme/orders — the URL SOLUTION.ts was already delivering to, which is how the mismatch surfaced — and the scorer counts only destinations pointed there. Matching is lenient on trailing slash, query string and case, strict on host and path.

Verified without an agent run

Using score-only, per the convention that a scorer change is validated for the price of some API calls:

• solution path: 2/2 pass
• --no-solution: fail
• solution edited to deliver to a receiver of its own choosing: fail — where it would previously have passed

Split out

The pnpm check breakage found while validating this is now #68, so the gate lands independently of a scenario change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MQzUoMAwEBJWpEGVvVzSjK