perf: inspect runtime overhead against sketch and raw exact computation

[zzylol]

Dependency stack: main → #768 → #737 → #749 → #771 → #728 → #770 → #763 → #765 → #761 → #742 → #759

Independent follow-ups to #765: #756 (diagnostics), #766 (runtime controls and overhead inspection).

Refs #758. Stacked directly on #765 as an independent tooling PR. It does not depend on #742, and #761/#759 do not depend on it.

Why and behavior

Before this PR: runtime workers followed implicit Tokio defaults, console/file logging could not be disabled independently, and there was no controlled comparison between direct computation, backend execution and HTTP query handling.

After this PR: --runtime-workers 2 --runtime-max-blocking-threads 8 --log-level off --disable-console-log --disable-file-log controls the production runtime and logging while retaining configuration artifacts. A Rust inspection executable scans server workers and client concurrency against direct DDSketch and raw exact median baselines, the real installed-plan query engine, and the real HTTP service. Every layer uses the same deterministic observations; forwarding is disabled and results must agree before measurements count as successful.

Reports retain p50/p95/p99 and samples, throughput, failed/dropped requests, process/driver CPU, RSS/peak RSS, observed threads, effective logging and visible cgroup limits. Fixed-rate arrivals include scheduling delay and count overload drops; closed-loop concurrency is also supported. Direct synchronous kernel timing is reported separately from task-dispatch latency. Each matrix cell uses a fresh process, with configurable warmup/repeats and JSON/CSV output.

The developer guide documents the shared runtime/task model, measurement boundaries, log comparisons and profiling commands. CI checks correctness and a small matrix, without imposing performance thresholds on shared runners.

Validation and limits

• Two runtime-control tests and four benchmark tests pass, including real backend/HTTP agreement with the direct sketch on one- and two-worker runtimes.
• A 16-cell functional matrix (four layers × two worker counts × two concurrency levels) completes 320 queries successfully; disabled sinks create no log files.
• Fixed-rate overload correctly records 1 completed / 99 dropped requests and passed=false. RUST_LOG precedence is verified.
• Production binary check, targeted Clippy with -D warnings, formatting and diff checks pass.
• Measurements above are debug-build functional validation, not performance results. No release speedup or eight-thread crossover is claimed. CPU profiling was not performed on this host (perf is unavailable).
• The initial fixture is one series, one complete pane and a DDSketch median. It excludes ongoing ingestion/maintenance and does not establish behavior for every sketch family or query shape. Memory includes the fixture, client and report buffers; process CPU minus driver CPU includes auxiliary threads.

Shared physical operator stack update

The shared runtime, operators and sketch codec are owned by ASAPPlanner #462, built on Planner #461. Backend #770 consumes them, #763 integrates ingestion and #765 integrates queries. The stack uses general semi-join and grouped Sort → Limit, with execution phase on the node. Independent library tests run in Planner; deployment acceptance stays in this backend stack.

Scope verification

The reorganized #759 and independent #766 trees pass strict workspace/all-target Clippy. Shared Rust runner tests pass. This does not resolve the existing #754 plan-shape gates or establish live performance benefit.