fix(data_plane): canonicalize sketch metric name at OTLP ingest

[zzylol]

Root cause

Warm-tier sketch queries (quantile_over_time/KLL, count/HLL,
topk/CountSketch) return status: error, "No result for query", data_source: asap_query even though the sketches are stored and the
controller's aggregations are registered.

The ASAPCollector edge pipeline's fused asapedgeprocessor
(opentelemetry-collector-contrib/processor/asapedgeprocessor/sketch.go:94)
sets MetricSuffix: "_" + family on every sketch it emits, so:

query metric (bare)	stored SketchInstanceMetadata.metric_name
request_size_bytes	request_size_bytes_kll
http_requests_total_latency_ms	http_requests_total_latency_ms_kll
unique_users_per_min	unique_users_per_min_hll
top_endpoint_qps	top_endpoint_qps_countsketch
endpoint_request_freq	endpoint_request_freq_countminsketch

The controller plans + pushes streaming-config policies keyed on the
bare metric name, and the query analyzer
(control_plane::asap_tier_analysis) lifts the bare name out of the
PromQL selector. With the suffix left on the stored name, both
resolution paths fail:
SketchIndex::instances_matching(bare, …) (equality on metric_name)
and find_matching_policies / find_policy_by_content (equality on
metric) return empty → EngineError::capability_miss. The 5-sketch
metrics aren't in backend-storage-routing.yaml so they're warm-only
(no archive fallthrough) → the adapter renders "No result for query".

http_requests_total (Sum family) is unaffected because the Sum path
(sum.go m.SetName(metricName)) does not suffix — which is
exactly why sum by (zone)(http_requests_total) works while the sketch
queries don't. Grouping (is_subset) is not the blocker:
per-series quantile_over_time has empty group_by_keys, which is a
subset of any stored sketch's keys.

Confirmed live on node2: GET /api/v1/db/schemas shows 20204 instances
under the suffixed names; GET /api/v1/streaming-config shows 6
policies under the bare names; with verbose RUST_LOG the engine logs
Modern execute() returned CapabilityMiss for each.

Fix

Per docs/design-controller-into-backend.md §1 the sketch family is
a wire-level attribute (already carried in agg_kind /
SketchKindHandle), NOT a name suffix. Canonicalize at the ingest seam
(data_plane/src/drivers/ingest/otel.rs):

• New canonical_sketch_metric_name(name, kind) strips the
  _<family> suffix only when it matches the datapoint's actual
  sketch kind (so a metric legitimately named foo_hll arriving as
  KLL is left untouched; never collapses to empty).
• Thread the canonical name through sid resolution
  (series_resolver.resolve), the series-key snapshot cache,
  SketchInstanceMetadata.metric_name, derive_sketch_policy_fp, and
  the legacy precompute-router match.
• The SeriesAssignment echoed back to the sender keeps the wire
  name (the exporter's cache contract is unchanged).

The strip is idempotent — a no-op once agents stop suffixing — so it is
safe alongside the eventual agent-side fix.

Remaining (agent-side, out of scope here)

topk(top_endpoint_qps) has a second mismatch: the agent emits a
heap-less CountSketch (indexed FrequencyEstimate) while topk(...)
needs a heap-bearing variant (FrequencyTopk). The real fix is the
edge processor emitting CountSketchWithHeap. After this PR + that agent
change, KLL/HLL/CountSketch queries all resolve warm.

Test plan

• cargo build -p data_plane green
• cargo test -p data_plane --lib — 718 passed (incl. 4 new
  canonical_metric_name_tests)
• cargo test -p control_plane --lib asap_tier_analysis — 57 passed
• Rebuild asap/data-plane:dev, redeploy on node2, soak ~90s, then
  quantile_over_time(0.99, request_size_bytes[30s]) /
  count(unique_users_per_min) return real data_source: asap_query
  results

🤖 Generated with Claude Code