feat(history): comparable runs and trustworthy regression alerts

[jeffbrennan]

Summary

Implements brief 06 — comparable history and trustworthy regression alerts, in three increments.

Increment 1 — versioned records & deterministic storage

• RunRecord v2 (RUN_RECORD_VERSION): provenance (status, backend, transport, compute_type, access_mode, runtime_version, client_version), workload_fingerprint, coverage, config; nullable measures; separate shuffle_read_bytes/shuffle_write_bytes and memory_bytes_spilled/disk_bytes_spilled; cumulative_time_s kept distinct from wall-clock duration_s.
• Legacy rows read as record_version=1 with unknown provenance and are excluded from baselines; historical zeros are never reinterpreted as measured.
• Deterministic format selection: .jsonl stays JSONL, existing Delta stays Delta; auto only uses the package heuristic for new stores.
• JSONL reads infer across the whole store so a long legacy prefix cannot hide v2 columns.

Increment 2 — baseline selection and rule assessments

• select_baseline_cohort requires the same workload identity (fingerprint, configurable), comparable coverage, and supports runtime/compute dimensions; filters null measurements before applying the window.
• check_alerts returns a status-bearing AlertAssessment per rule (triggered/clean/insufficient_data/unsupported/skipped), with configurable min_samples, explicit true-zero handling, and finite-threshold/positive-window validation.
• Capture forwards alert_output_path for on_trigger="file" and exposes alert_assessments.

Increment 3 — comparison report

• workload_fingerprint (stable across node ids, changes with join type/scan path).
• compare_runs + sparkparse compare CLI: compatible metrics, cohort/sample counts, excluded measures with reasons, and plan_changed.

Validation

• just ci — ruff, ruff format, pyrefly (0 errors), 477 passed / 2 skipped.
• Reviews addressed: cumulative time deduplicated per query; mixed-version JSONL schema inference; window counts only valid measurements.