feat(telemetry): every sample of one MCP tool call carries its operation id (#733, slice 3) - #967

Summary

Issue #733, slice 3 (Refs #733, the issue stays open). Root cause: the samples one MCP tool call causes were joined only by op and ts (docs/audits/operation-consumption-inventory-2026-11-05.md, "No operation identity"). A nested handler's sample, a worker thread's sample and a sample recorded by work that outlives the call had nothing pointing to the call.

telemetry.capture_records() (opened by TelemetryMiddleware for every tools/call) now sets a fresh operation_id (uuid4 hex) in a ContextVar. record() stamps it on each TelemetrySample. Every sample recorded in that context carries the same id, and so does every sample recorded in a context copied from it: a run_in_executor worker or a task that records after the call has closed. A sample recorded outside a tool call (hooks, capture worker) carries operation_id: null, an explicit "not attributed", not an invented id. The field reaches telemetry.jsonl and the exporter's sample dict. The OTLP exporter reads only op and the outcome, so the id never becomes a metric attribute (unbounded cardinality, written in the TelemetryExporter contract).

record() delegates its capture-or-publish tail to _deliver() so it stays under the 40-line method rule.

Refs #733

Type of change

• New feature (non-breaking change which adds functionality)

Test plan

• New: tests_py/core/test_operation_identity_733.py (4 tests): nested and outer samples share one id; two calls get distinct ids; a worker-thread sample and a sample recorded after the call keep the caller's id; a sample outside a call has None. Removing the operation_id stamp, or resetting it per sample, fails them.
• pytest tests_py/core tests_py/test_telemetry_middleware.py tests_py/handlers tests_py/hooks: 6735 passed, 276 skipped, 3 failed. All 3 fail the same way on main in this container: test_wiki_coverage_dashboard.py::test_unreadable_page_is_skipped_not_fatal (runs as root, so the file is still readable) and 2 cases of test_entry_auto_dispatch.py.
• pytest tests_py/infrastructure tests_py/*.py tests_py/observability tests_py/server tests_py/invariants tests_py/architecture: the only failures are 4 sentence_transformers/torch tests. torch is not installed here because the proxy refuses download-r2.pytorch.org. This diff does not touch that code.
• ruff check and ruff format --check (0.17.3, the version in uv.lock) on the changed files: clean. python scripts/check_craftsmanship.py: OK.

Audit notes

• Not verified: pyright run locally; PostgreSQL paths (no server here; telemetry does not touch the store).
• Remaining slices of feat: expose local operation receipts with attributable resource consumption #733: background capture (capture_processed, resident worker) does not run in the tool call's context, so it still carries null and needs the id passed explicitly through the spool; retry count; CPU time; provider usage beyond total_cost_usd.

Coding-standards compliance

• §2.2 Layer dependency direction preserved (core uses only stdlib uuid, as core/wiki_identity.py does).
• §4.1 / §4.2 300 / 40 limits (craftsmanship gate OK).
• §9 No dead code.

Breaking changes

None. telemetry.jsonl lines gain one key, operation_id.

Generated by Claude Code