fix(otlp): preserve empty metric descriptors under non-matching policies

[detail-app]

Detail bug report: View on Detail

Fixes ENG-704

Bug

When any metric policy was loaded (even one matching nothing), the OTLP proxy silently dropped metric descriptors whose data oneof was set but whose data_points list was empty — a valid OTLP wire form (a named metric with type/unit/description but no measurements this batch). dropped_count stayed 0, so the loss was invisible to monitoring.

Two compounding defects in src/signals/otlp/metrics.zig:

• filterMetricsInPlace pruned any metric with data != null and dataPointCount == 0, but checked only the post-filter count — it could not distinguish a metric that arrived empty (a valid descriptor) from one emptied by policy. This contradicted the code's own comment ("A metric with no data at all is left alone").
• processProtobufMetrics always re-encoded (no "nothing-modified" fast path), so the prune ran even when no policy matched — unlike processProtobufLogs, which returns the original bytes verbatim when nothing was modified.

Fix

• Track the pre-filter count in filterMetricsInPlace: capture original_dp = dataPointCount(metric) before per-data-point filtering, and prune only when original_dp > 0 and dataPointCount == 0 — i.e. only metrics emptied by policy are pruned; originally-empty descriptors survive. This fix is shared by the protobuf and JSON paths.
• Add the nothing-modified fast path to processProtobufMetrics, mirroring processProtobufLogs: when counts.dropped_count == 0, return the original bytes without re-encoding. This restores logs/metrics parity and prevents the prune from firing for non-matching policies.

Testing

• Unit tests (zig build test --summary all, Debug and ReleaseSafe): 522/523 pass, 1 skip (the pre-existing S3/MinIO e2e test that self-skips via return error.SkipZigTest; unrelated to this change). Added 3 regression tests:
  • empty descriptor survives a non-matching policy on the protobuf path,
  • a policy-emptied metric is still pruned while an originally-empty metric alongside it survives (the key invariant of the prune fix),
  • empty descriptor survives a non-matching policy on the JSON path.
• Negative verification: reverting only the prune condition (keeping the fast path) makes the partial-filter test and the JSON repro fail by name, confirming the prune fix is load-bearing.
• Build / lint / format: zig build, zig build -Doptimize=ReleaseSafe, zig fmt --check src/, and ziglint all pass clean.
• End-to-end: built edge-otlp and the repo's echo-server (as a live upstream), loaded a non-matching metric policy, and generated valid OTLP payloads with protobuf + opentelemetry-proto. The upstream captured both the populated gauge and the empty-gauge descriptor (53 bytes, byte-identical to input — confirming the streaming path returns .keep, not .replace). Repeating with a policy that drops one metric by name confirmed the emptied metric was pruned while the originally-empty descriptor survived. As a final check, rebuilding edge-otlp with the fix reverted reproduced the bug end-to-end (upstream received only the populated gauge, 36 bytes), then the restored fix forwarded both metrics again.

---

Automatic Fixes PRs can be configured here.

[macroscopeapp]

Approvability

Verdict: Approved at 809ec56

Macroscope's review found this PR approvable — This is a narrowly scoped OTLP metrics bug fix with regression coverage for both protobuf and JSON paths. It preserves valid empty descriptors while retaining existing policy-driven pruning, without introducing schema, infrastructure, or sensitive-data changes.

^{You can add or adjust custom eligibility rules. Learn more.}