feat(sketchdb): cold-query fallback with local-FS raw store (§5.2)

[zzylol]

Addresses TODO.md blocker #1 — the paper's "sketches for hot, exact for cold" story needs a cold path that serves capability-miss queries from raw samples instead of hitting the live Prometheus (or failing).

Design

• New ColdStore trait + LocalFsColdStore impl (v1). On-disk layout is byte-identical to the future S3 key layout:
  ```
  /raw//YYYY/MM/DD/HH/part-NNNNNN.jsonl
  ```
  Each line: `{"ts_ms": int64, "labels": {...}, "value": f64}`. Swapping to S3 is later a new impl of `ColdStore` — no adapter changes.
• New `ColdFallback<S: ColdStore>` implements `FallbackClient`. Parses PromQL, extracts `(metric, predicates, op)`, scans the cold store, computes the answer. Supported shapes:
  • bare instant vector `metric{labels}`
  • no-grouping scalar aggregations (`sum|count|avg|min|max`)
  • matchers: `=`, `!=`
• Unsupported shapes (regex matchers, `rate()`, `by (...)`, binary ops, etc.) delegate to the optional inner `FallbackClient` (chain-of-responsibility — typically the existing Prometheus proxy).
• Hot / cold telemetry counters (`queryengine_hot_queries_total`, `queryengine_cold_queries_total`, `queryengine_cold_bytes_served_total`), keyed by `(metric, shape)`. Mirror the PR feat(metrics): wire §6.3 barrier counter into all OTLP ingest paths #51 pattern.
• `AdapterConfig::prometheus_promql_with_cold` convenience constructor.

Routing

No engine changes. `process_query_request` already falls through to `FallbackClient` on engine-miss; configuring a `ColdFallback` naturally routes Purged / capability-miss queries through the cold tier, with the Prometheus proxy as tail-of-chain for shapes the cold adapter doesn't recognise.

Tests

752 → 777 (+25):

• 5 unit tests on format + hour-prefix helpers
• 4 on `LocalFsColdStore::scan` (range filter, missing prefix, hour-boundary span)
• 10 on `ColdFallback` plan extraction + aggregation + float formatting
• 6 end-to-end HTTP integration tests in `tests/cold_fallback_tests.rs` (bare selector, sum aggregation, label filtering, unsupported-shape delegation, telemetry increments, and the canonical "Purged range served from raw samples" paper claim).

clippy clean, fmt clean. One pre-existing flaky elastic forwarding test is unrelated (passes in isolation).

🤖 Generated with Claude Code