feat(analyze): correct metrics and evidence-based findings

[jeffbrennan]

Implements brief 03, all three increments.

Metric contract

• New sparkparse/metrics.py: a canonical registry mapping verified classic and Photon/Connect raw metric names onto canonical names with unit, scope, aggregation, coverage and derivation. Unmapped metrics are preserved with canonical: null rather than guessed — numRowsScanned ≠ numOutputRows, data size is not scan bytes, operator spill is not task memory/disk spill, and Photon's numBytesRead (shuffle on an exchange, files on a scan) stays unmapped.
• Exact integer counts survive end to end: accumulator_totals now carries value_exact (Int64) beside the Float64 value in both ingestion paths, populated for every metric that is not rescaled. A count of 2**53 + 1 makes it from the event log or Connect server through to the JSON export.
• A missing metric is null, a measured zero is 0. node_duration_minutes is now null when a node reported no timing metric, instead of claiming it ran instantly.
• Empty task frames yield null totals, not zeros.

Rules

• find_cartesian_joins means cartesian: CartesianProduct, nested loops with no condition, or join_type=Cross, matched per (query_id, node_id). Conditional nested loops moved to find_nested_loop_joins — the nested_loop_join fixture's BNLJ is LeftOuter with a condition and is no longer reported as cartesian.
• Row expansion is computed from a join's immediate inputs, descending only through row-preserving operators. Scans identify lineage only.
• Repeated scans count occurrences, so repeats inside one query are visible, and track distinct read schemas.
• find_largest_scans labels bytes_source and falls back to rows; nothing observed means nothing is presented as "largest".
• New: scan efficiency, task stragglers (promoted to data skew only with size evidence), GC overhead, shuffle volume (sides reported separately), AQE runtime adjustments.
• Every rule returns a RuleAssessment (evaluated / unsupported / insufficient_data) and checks its counters are actually populated before reporting evaluated, so a null column can't read as a confident zero.

Presentation

• Findings carry evidence with units, thresholds, confidence, caveats and a next investigation, and stay separate from the raw plan summary.
• sparkparse analyze gains --compact, --top-n, --redact, --findings. Redaction covers node names, which embed the server's operator text for unmapped Connect operators.
• The dashboard shows confidence and caveats plus a "Checks not run" panel, so missing telemetry is explained rather than hidden.

Known limits, stated rather than papered over

• ScanDetail does not retain PushedFilters/PartitionFilters, so scan efficiency reports rows read but discarded — not a pruning verdict.
• parse.py keeps only isFinalPlan=true snapshots, so aqe_plan_change reports optimizer-recorded adjustments on the surviving plan and says so in its assessment reason.

Validation

Offline only — no live Spark or Databricks run. just ci: 377 passed, 2 skipped; ruff clean; pyrefly unchanged at the two pre-existing errors (pages/home.py:132, pages/summary.py:172).

New tests: tests/test_metrics.py, synthetic-fixture and Photon-fixture coverage in tests/test_analyze.py (reused node IDs across queries, CartesianProduct, conditional nested loops, zero-row inputs, filters before joins, reused exchange, repeated scan within one query, incomplete scan detail, plan spill without task rows, all-null scan ranking, unsupported/insufficient statuses, redaction of an unknown operator carrying a literal, and ingestion-to-JSON precision above 2**53). Helpers live in tests/synthetic.py.

Expected-output fixtures were regenerated; the only field changes are node_duration_minutes (0.0 → null) and the new value_exact.