fix: endpoint failures use the standard error shape and count in metrics

[blaipr]

What

• The output endpoint's failure paths (500 script/builtin failure, 504 timeout, the neither-output-nor-script guard) rendered hand-built json! bodies; they now flow through ApiError — one source for the shape, finishing the migration error.rs describes.
• Metrics bug: a timed-out run returned before the counters, so unified_api_endpoint_total never counted it — despite docs/endpoints.md promising that timed-out runs count as result="error" ("so alerting on the error rate catches hung transformers too"). Timeouts and builtin-render panics now flow through the counters and the duration histogram before answering.
• The 503 keeps its missing_sources list (now sorted, deterministic) under a declared EndpointUnavailableBody schema instead of an undeclared ad-hoc body.
• #[utoipa::path] on both endpoint handlers now declares 500/503/504 with their bodies, and GET /metrics joins the OpenAPI spec — it was the only routed handler missing from it, and the spec never mentioned that server.metrics_require_auth moves it behind the key.

Tests

New tests/adapters/out/output/slow.py sample transformer plus an integration test: a script endpoint with timeout_seconds: 1 answers 504 whose body is the standard {"error": ...} shape naming the limit it hit.

Note: a_cron_scheduled_source_syncs_into_the_cache and a_reload_replaces_the_running_task_generation flaked once under full-suite parallel load and pass in isolation and on re-run — pre-existing timing sensitivity, unrelated to this diff.