feat(promql): type-driven histogram_quantile discrimination (#79)

[zzylol]

Closes #79.

Problem

histogram_quantile(φ, m) picks between exact classic-bucket interpolation (HistogramQuantile, not sketch-able) and the generic sketch-able Quantile using a purely structural heuristic (by (le) grouping / _bucket metric name / le= matcher). The true signal is the argument's sample type, which structure only proxies — so the heuristic has real failure modes:

• false positive — a metric merely named …_bucket (or a native histogram so named) → interpolation;
• false negative — a classic histogram exposed without _bucket and queried without le → the sketch path, which is wrong (you can't sketch pre-aggregated buckets).

Fix: drive it from declared sample type

• histogram module: HistogramKind (ClassicBucket / Native / RawSamples — only ClassicBucket is non-sketch-able) and a HistogramCatalog (metric → kind) the client supplies.
• histogram_arg_is_sketchable: a declared kind for any metric referenced in the argument decides it; otherwise fall back to the existing structural heuristic. So metadata overrides structure where the sample type is known, and undeclared metrics behave exactly as before.
• New entry point lower_promql_with_histograms(query, accuracy, catalog). lower_promql stays heuristic-only.

This directly serves the design the two lowerings exist for: a client holding raw samples can declare them RawSamples and get the sketch-able Quantile even when the user wrote histogram_quantile.

On the plumbing

The discrimination is consulted in exactly one place (walk_histogram), reached from deep in the recursive, free-function walk chain. Rather than thread a catalog parameter through ~20 functions / 36 call sites, the catalog is installed as an RAII-guarded thread-local for the duration of one lowering call (lowering is synchronous, one query at a time). A test asserts it doesn't leak across calls.

Scope

The heuristic is kept as the deliberate fallback — not every metric will carry metadata, so removing it would regress undeclared queries. #79 also floated per-series tagging and an L4-deferred variant; per-metric catalog is the right granularity for the lowering-time decision and is what the client can actually declare. The mechanism is now in place for a production catalog to feed.

Tests

histogram_metadata.rs:

• heuristic baseline unchanged without a catalog;
• false-negative fix: declared ClassicBucket on a suffix-less metric → HistogramQuantile;
• false-positive fix: declared RawSamples/Native on a _bucket-named metric → Quantile;
• undeclared metric falls back to the heuristic;
• the ambient catalog does not leak across calls.

Unit tests cover is_sketchable, catalog lookup, and guard install/restore. Full workspace suite green; clippy clean.