bench: measure ClickHouse and MetricsQL o11y paths

[zzylol]

Why

The ClickHouse SQL and VictoriaMetrics MetricsQL paths need an auditable experiment that keeps all 27 observability queries and every fallback in the denominator. The experiment must reject a benefit claim when a language frontend, published plan, or cross-engine semantic comparison fails.

What

This PR adds a deterministic fixture generator, physical-plan generator, production frontend/compiler auditors, one-command fresh-service runner, structured result comparator, per-query terminal-stage report, and checked-in evidence for one fresh trial with three repetitions.

How

reproduce.py starts from an empty output directory. It generates the OpenMetrics fixture, Prometheus config and TSDB, and physical plan; records their hashes; builds the data-plane binary from that same clean HEAD; runs the production SQL/MetricsQL frontend and compiler paths; starts fresh Prometheus, VictoriaMetrics, ClickHouse, and data-plane state; randomizes query order with a recorded seed; alternates engine order; and records phase-specific mixed-phase CPU ticks, RSS/HWM, storage, ingest time, lifecycle time, raw responses, and structured comparisons.

Before this PR

There was no self-contained way to reproduce the 27-query multi-language measurement. Earlier evidence depended on undeclared /tmp state, recorded a stale backend identity, did not structure all protocol comparisons, and overstated the scope of the observed trials.

After this PR

One command regenerates the checked evidence without a pre-existing TSDB or fixture. The observed result is deliberately narrow and does not establish acceleration benefit:

• One fresh trial, three repetitions, 27 queries: 81/81 responses from each engine.
• ASAP MetricsQL: 81/81 exact_fallback; median/p95 5.03/6.34 ms. Native VM: 2.36/9.12 ms.
• ASAP ClickHouse: 81/81 exact_fallback; median/p95 56.25/179.34 ms. Native ClickHouse: 49.92/169.56 ms.
• ClickHouse versus ASAP ClickHouse: 81/81 structured matches.
• VictoriaMetrics versus ASAP MetricsQL: 81/81 structured matches.
• Prometheus versus VictoriaMetrics: 48/81 strict matches and 33/81 mismatches.
• Prometheus versus ClickHouse SQL label/value oracle: 81/81 matches. The corpus now records oracle_valid for 27/27 mappings.
• MetricsQL parser 12/27, early planner 7/27, production compiler/publication 2/27. SQL acceleration frontend 0/27.

Evidence

tools/o11y-multilang/repro-fresh/manifest.json records clean source HEAD 806945b3, the binary SHA-256, immutable image digests, generated fixture and plan hashes, lifecycle/ingest/query duration, and start/post-ingest/post-query resource snapshots. Raw and structured per-request evidence is checked in alongside it.

Verification

• Python syntax: python3 -m py_compile tools/o11y-multilang/*.py — passed.
• Rust formatting: rustfmt --edition 2021 --check over all three benchmark examples — passed.
• Rust examples: cargo check -p control_plane --examples with the shared target directory — passed.
• Reproduction: reproduce.py --trials 1 --repetitions 3 --seed 20260910 from an empty output directory — passed.
• Structured comparison: 291 matches, 33 Prometheus/VM mismatches, zero comparison errors across the four comparison pairs.
• Repository checks: git diff --check passed; generated state directories, logs, TSDB blocks, credentials, and private tokens are absent from the diff.

Architectural decisions

The benchmark uses the production language frontends and compiler APIs and leaves PromQL/SDS contracts unchanged. Exact fallback is a terminal measured outcome. Cross-protocol Prometheus/ClickHouse comparison treats labels and values as semantic fields and reports timestamps and result type as non-comparable; native/proxy pairs compare labels, values, timestamps, result type, and warnings.

Limitations and follow-up

This is one fresh trial, so it does not estimate between-trial variance. CPU is process scheduler ticks rather than normalized CPU time, and process totals cover mixed alternating modes, and container writable-layer size is an operational storage proxy. The generated physical plan contains no corpus sidecars: even q03 and q16 pass offline publication validation; their catalog miss is inferred from the sidecar-free measured artifact, while exact fallback is observed on HTTP. Binder, validator, executor, and warm performance are not exercised by the 27-query run. Native histogram exponential interpolation is outside the classic _bucket q21 query.

Human review — do not complete with an agent

• The MVP boundary is correct.
• New conceptual layers or public interfaces are necessary.
• The before/after description matches the intended product behavior.
• Human reviewer:
• Decision and rationale:

[zzylol]

Closing because this benchmark is based on the obsolete #559 architecture and its checked-in run measured 27/27 exact fallback before the unified ClickHouse path reached main. A fresh mixed differential and 27-query execution matrix must be generated from current main.