fix(worker): install shutdown signals before readiness; test(engine): isolate OTLP ingestion metrics storage - #2235

Split out of #2198 (which is now scoped to #2178 only), as requested in review, to ease review and bisecting. Neither change touches trigger handling.

Changes

• fix(worker): install the shutdown signal handlers before exit-watch reports readiness (crates/iii-worker/src/daemon_exit.rs), so a signal sent right after readiness is not missed. Covered by sandbox_daemon_exit_integration.
• test(engine): OTLP ingestion edge-case tests now use owned metric storage through an extracted ingest_otlp_metrics_into(json, storage) helper (otel.rs), instead of clearing the process-wide store. This fixes a CI flake (test_ingest_metrics_multiple_resources counted 3 metrics instead of 2) caused by a non-serialized test writing to the same global store. Production behaviour is unchanged: ingest_otlp_metrics still ingests into the global store.
  • The misleading "interleaved writer" test (a thread spawned and joined immediately) was rewritten as a sequential ..._ignores_foreign_writer test.

Validation

• cargo fmt --check ok
• cargo test -p iii --lib workers::observability: 442 passed
• cargo test -p iii-worker daemon_exit: 5 passed + 1 integration passed

Summary by CodeRabbit

• Bug Fixes
  • Improved shutdown-signal handling so signals are registered before exit monitoring begins. If one signal handler fails or its stream closes, other shutdown and engine-exit checks remain active.
  • Metric ingestion now handles unavailable storage without attempting to store metrics, while malformed input continues to return an error. Empty batches and malformed batches are also covered by additional checks.