test(analyzer): α — 18-query parity catalog for analyzer unification

[zzylol]

Summary

PR α of the analyzer-unification chain (α→β→γ→δ→ε). Pure additive
baseline that freezes what each PromQL → warm-tier analyzer answers
today, so β/γ have a verifiable parity contract before they start
collapsing the two surfaces.

• analyzer_parity_tests inline #[cfg(test)] mod in
  data_plane/src/query_engines/asap_query_engine/engine.rs: 18-query
  corpus covering quantile_over_time, sum-by, topk,
  count_over_time, histogram_quantile, bare selectors,
  rate/irate/increase, sum(rate(...)), and unparseable input.
  Single assert_eq! against an embedded golden-master captured on
  origin/main@6557fb8.
• controller/docs/analyzer-parity-matrix.md (168 lines): human
  catalog of the same 18 rows. Documents the 5 controller-rejects-but-
  engine-accepts divergences the chain is meant to resolve — D1
  sum_by(m), D2 topk(k, m), D3 count_over_time(m[r]),
  D4 rate/increase, D5 sum(rate(m[r])) — plus β/γ/δ/ε
  amendment rules and pointers to both pipelines' source files.

Findings worth flagging (caught by the corpus)

• The TODO doc's claim "engine accepts irate" is outdated: irate
  is not in the engine's pattern list — both analyzers reject it
  (parity).
• The γ5 / PR refactor(legacy_expr): step γ5 — histogram_quantile is parser-level substitution #144 histogram_quantile parser substitution only
  handles histogram_quantile(phi, MatrixSelector); the composed
  Prometheus-canonical histogram_quantile(phi, sum_by(le, rate(m[r])))
  is rejected with UnparseableMetricsql("expected MatrixSelector, got Discriminant(...)"). Worth a re-examination when ε expands
  the corpus.
• Post-PR refactor: consolidate Capability into sketch_algebra; warm_tier_analysis becomes facade (Step 2a) #129 (Step 2a) the controller returns
  Capability::QuantileApprox(Any), not QuantileApprox(DDSketch)
  as the TODO doc references. Family-wildcard form; downstream
  is_satisfied_by does the family match.

Scope / risk

• 255 LOC additive in engine.rs, 168 LOC new doc. Zero production
  code edits, zero visibility changes, zero new deps.
• parse_and_match_promql and build_query_requirements_promql stay
  private — the test lives in the same module and reaches them via
  super::*.
• Branch built off origin/main@6557fb8 so it rebases cleanly onto
  the in-flight feat/schema-retire-final-* slices.

Test plan

• cargo build -p data_plane --lib clean (6 pre-existing warnings)
• cargo test -p data_plane --lib analyzer_parity_tests → 1 passed
• cargo test -p controller --lib warm_tier_analysis → 19 passed (unaffected)
• Reviewer: open controller/docs/analyzer-parity-matrix.md
  alongside the golden in engine.rs and confirm each row's
  D{1..5} annotation matches your expectations before β starts.
• Reviewer: confirm the chosen 18 queries cover the shapes you
  want γ to preserve. Anything to add now (cheap) vs in ε
  (expensive — diff churn against the embedded golden)?

🤖 Generated with Claude Code