Export the Neutron error-kind breakdown as a dedicated OTLP counter

[berendt]

Closes #25

Summary

The OTLP export from #19 collapses every failure to outcome ∈ {success, error, timeout} on openstack_tester.operation.duration, so a metrics backend can see that Neutron calls fail but not why. This branch exports the neutron client's errKind classification as a dedicated counter, openstack_tester.operation.errors, so a PromQL backend can answer "which kinds of errors, on which resource kind and operation, at what rate" — the run report's Errors table, over time. The five existing instrument families keep their exact schema; the counter is a sixth instrument that fires only on failures.

Change set (in commit order)

a963904 — Export failed Neutron operations as an errors-by-kind counter

• Adds the sixth OTLP instrument in internal/telemetry/telemetry.go: the monotonic counter openstack_tester.operation.errors with attributes kind, operation, and error.kind (Prometheus: openstack_tester_operation_errors_total, label error_kind).
• Recorded from Telemetry.RecordOperation whenever errKind != "". The seam already receives the neutron client's verbatim errKind, so no call sites in internal/neutron change and the existing five instrument families are byte-identical in schema.
• A counter rather than an error.kind attribute on the histogram: it adds one cheap series per (kind, operation, error_kind) that actually occurs, only on failures, instead of multiplying the histogram's ~15-series streams by the error-kind set — and it leaves the documented operation.duration schema untouched.
• Tests: telemetry_test.go asserts the emitted schema via the ManualReader provider (a failing op yields the counter with the expected attributes; a success yields nothing); neutron/telemetry_test.go covers the classification end-to-end including a quota, an http_<status>, and a success case.

8be04c7 — Document the operation errors counter in README and otel-verify

• Adds the openstack_tester.operation.errors row to README §9's instrument table, documents error.kind's bounded value set (quota, timeout, canceled, other, http_<status>) and its Prometheus surface, notes it is recorded only for failed operations, and adds the by-error-kind cookbook query.
• Records — in the Makefile comment and README — that this family is deliberately absent from the make otel-verify required set: a healthy run produces no errors, so its absence is the expected steady state, not a missing-metric failure.

65e6e29 — Add an errors-by-kind panel to the API operations dashboard

• Adds a full-width "Errors by error kind" stacked timeseries to the Neutron API operations dashboard (contrib/otel/dashboards/api-operations.json), driven by openstack_tester_operation_errors_total summed by kind, operation, and error_kind — the report's Errors table over time, empty on a healthy run.
• Retires the "error kinds collapse to outcome by design" forward-pointers in the api-operations and overview dashboard descriptions and in README §9, now that the counter and its panel exist.

Divergence from the issue

None material. The issue anticipated (step 5) that "whichever issue lands second picks up" the dashboard breakdown; #24's dashboards landed first, so 65e6e29 supplies the panel here. All four acceptance criteria are covered: the new counter fires with the right attributes on failure and not on success, the existing five families are schema-identical, README §9 documents the instrument and cookbook query, and unit tests cover quota, http_<status>, and success cases.

---

Implemented by planwerk-agent 6b3f3ea with Claude:claude-opus-4-8