`.res-body` is display:block, so `.chart-view { flex: 1 }` was inert: the view
collapsed to the config bar's height and Chart.js sized the canvas to a few
pixels — a config bar with no visible chart. Use `height: 100%` (the block
parent has a definite height) for `.chart-view` and `.chart-empty` so the
canvas wrapper gets real height.

Not caught earlier: happy-dom does no layout, and the manual browser harness
attached the canvas into a pre-sized container, bypassing the real `.res-body`
parent. Verified live against an otel system.query_log result.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Wtzg34E1iGzKm11TArRpko

Vitest 2.x defaults to `pool: 'forks'`, which fans out to (cpus-1) child
node processes via tinypool — 11 on a 12-core box. On normal exit those
should be reaped, but a detached swarm can survive a run and accumulate
across runs until it saturates CPU/RAM.

Switch to `pool: 'threads'` (capped at 4): worker threads live inside the
single vitest process, so they die with the parent and can never become
orphaned OS processes. Safe here — the suite is pure ES modules + happy-dom
with no native deps. Also faster (happy-dom env setup 21.7s -> 3.0s, total
6.1s -> 2.3s) since threads skip per-fork environment spin-up.

Belt-and-suspenders: a `pretest` hook kills any pre-existing stray vitest
processes before a run.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016SGQKuUW7a12hBSwmwSkYa

…bugs

Persistence
- Save/restore the per-tab chart config (type/X/Y/multi/series) with saved
  queries and share links. Saved entries gain an optional { cfg, key }; share
  links carry a tagged JSON envelope (back-compatible with legacy SQL-only
  hashes); export/import round-trip it. Configs are deep-cloned on save/restore
  so a restored chart can't mutate its source, and a chartCfgValid guard rejects
  out-of-range indices from a hand-edited link/import.

Chart-view fixes
- chartLabel keeps day granularity (YYYY-MM-DD, display-only); buildChartData
  SUM-aggregates per cell instead of last-wins, so duplicate buckets combine.
- ORDINAL_RE anchored (^bucketword s?$) so measures like monthly_revenue are no
  longer misread as ordinal axes.
- running guard runs before chartCfgFor, so a still-settling result can't stamp
  over a restored config.
- normalizeChartCfg clears a Series equal to X and folds pie to single-measure /
  no-series; chartCfgFor normalizes on every restore.
- renderChart plots rows in query order, independent of the table sort (also
  drops the wasteful full-result sort before the 500-row cap).
- chartColors resolves --mono to a real font stack for canvas ticks/legend.
- .chart-view scrolls and the canvas keeps a min height instead of crushing to 0.
- pretest pkill scoped to this project's vitest config.

523 tests pass; pure/render layers at 100/100/100/100. Verified e2e on antalya
against ontime.fact_ontime.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016SGQKuUW7a12hBSwmwSkYa

…ay labels, scope pretest

- chartCfgFor is now the single place that folds chart-config invariants; the
  Type/X/Series handlers just mutate + re-render (drop their redundant
  normalizeChartCfg calls).
- chartLabel keeps date+HH:MM for an intraday timestamp (date-only for a Date or
  midnight DateTime), so two same-day buckets no longer collapse to the same
  tick text. Grouping is still on the raw value. Removed the now-unused
  ISO_DATE_RE.
- pretest pkill pattern requires both `vitest` and `--config tests/vitest.config.ts`
  so it can't match an unrelated process that merely has the path in its argv.

525 tests pass; chart-data.js at 100/100/100/100.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016SGQKuUW7a12hBSwmwSkYa

- Saved queries now persist the result view (Table/JSON/Chart) alongside the
  chart config; saveQuery captures state.resultView (the transient raw view is
  not saved), and saved-io export/import/merge carry it.
- Opening a saved query reopens that view and runs immediately: the saved-row
  click calls run({ view }), and run() now starts in the given view (or Table
  when absent/unknown) instead of always resetting to Table. History rows also
  re-run on click (Table view).
- View ownership lives in run() (which already resets resultView); loadIntoNewTab
  stays responsible only for tab-level state (sql/savedId/chart).

528 tests pass; core/state/render layers at their per-file gates. Verified e2e
on antalya: a one-click saved-Chart query auto-runs straight into its chart.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016SGQKuUW7a12hBSwmwSkYa

…143)

`pkill -f` matches full command lines, so the hook's own shell — whose argv
contains the pattern string — was a match, and pkill SIGTERM'd its parent
(`npm test`) before `|| true` could run. CI fails at the pretest step with exit
143 before vitest starts; it only passed locally because I'd been running the
vitest binary directly, bypassing the npm lifecycle.

The hook guarded against the forks-pool orphan swarm, which `pool: 'threads'`
already eliminated (threads die with the parent — verified zero lingering
processes), so it's removed rather than patched.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016SGQKuUW7a12hBSwmwSkYa