feat(ingest): unified registry-allocated sid (PR-1+2+3, precompute follow-up in PR-4)

[zzylol]

Summary

Three-commit chain implementing Option B (backend-allocated registry) for sid mint on the OTel ingest path. Sid identity is (metric, attrs_fingerprint, agg_kind_canonical) — the same 3-tuple compute_sketch_sid hashed over, just held as a registry-allocated u64 instead of content-addressed.

Why registry (B) over u128 widening (A): xxh64 gives only probabilistic uniqueness; the wire shape requires uniqueness as a contract (agent omits attrs on subsequent emits → receiver must disambiguate sid → (metric, attrs, agg_kind)). Registry approach gives uniqueness by construction (AtomicU64::fetch_add). Trade-off accepted: sids are not stable across independent backends; cross-restart durability solved by WAL persistence.

Commits

• 84abac7 — PR-1: SeriesIdResolver authoritative on ingest path. Replaces compute_sketch_sid call at otel.rs:880 with series_resolver.resolve. Returns IngestOutcome { unknown_series_ids, series_assignments } so the gRPC Export response can echo the canonical sid back.
• 3ac90cf — PR-2: WAL-backed persistence. SeriesResolverPersistence trait + NoopPersistence + FilePersistence. WAL append-only, single-writer, fsync per record. SeriesIdResolver::open(path) replays on construction, resumes next_sid at max(replayed)+1. Torn-write detection truncates to last safe offset. Wired in main.rs under --persistence-enabled.
• f72b533 — PR-3: identity = (metric, fp, agg_kind); delete compute_sketch_sid. Fixes a latent PR-1 regression (two sketch kinds over same series collapsed to one sid). WAL bumped to v2 (4-field records). compute_sketch_sid + 5 unit tests deleted.

Design contract pinned

sid = registry-allocated u64 for (metric, attrs_fingerprint, agg_kind_canonical)

• One mint authority: SeriesIdResolver.
• Bandwidth optimization: sender caches sid via series_assignments reply, omits attrs on subsequent emits.
• Recovery: stale sid → unknown_series_ids reply → sender evicts + re-emits with attrs → resolver hits cache or mints fresh.
• Durability: WAL replay preserves (metric, fp, agg_kind) → sid across restart.

WAL format v2

header: 8 bytes  →  b"ASAPSRP\x02"
record: u64 sid LE
        u32 metric_len LE  + metric utf8
        u32 fp_len LE      + fp utf8
        u32 agg_kind_len LE + agg_kind_canonical utf8

Caps: metric ≤16KiB, fp ≤64KiB, agg_kind ≤4KiB. No CRC for v2; add if telemetry surfaces bit-rot.

ResolveSeriesIDs RPC stubbed

The pre-resolve gRPC handshake at otel.rs::resolve_series_i_ds returns empty assignments and logs a one-shot WARN. The proto's SeriesQuery doesn't carry agg_kind — a pre-resolve there can't produce the right sid under the new identity model. Agents that ignored the RPC and just sent Export-with-attrs work unchanged. Follow-up: drop the RPC, or extend SeriesQuery to carry agg_kind_canonical.

PR-4 (separate follow-up)

compute_sid is not deleted here — still called by SketchStore::ingest_precompute_for_agg_config for PRECOMPUTE aggregations (5 call sites: output_sink, eviction service, backfill processor, two test sites). PR-4 migrates those to the resolver and deletes compute_sid + sketch_kind_tag + encode_sketch_config. Closes the inconsistency — one mint authority for every sid in the system, matching the stated goal of "single sid across asapcollector and asapquery-backend; precompute is just an extension of the OTel processors at the edge."

Test plan

• cargo build -p data_plane (lib + bin) — clean (5 pre-existing warnings)
• cargo test -p data_plane --lib — 763/763 passing
• cargo test -p data_plane --lib drivers::ingest — 24/24 passing (was 14 pre-chain; +8 persistence + +2 identity tests)
• Reviewer: confirm the wire-case comment block at otel.rs:867 pins the right design contract (Interpretation B — sid = (metric, fp, agg_kind)).
• Operational: WAL grows ~250-400 bytes/record. At 100M sids ≈ 25-40 GB. Compaction is out of scope; revisit if cardinality climbs.
• Patched OTel-Go exporter: nothing required. Export path's series_assignments reply is what the exporter caches today; ResolveSeriesIDs was optional and now returns empty (graceful degradation).