…ine)

Snapshot of in-progress async-worker refactor before reworking swap-axis to
move all per-setting computation (incl. grouping) into the worker. Includes
new ui/src/lib/swap.ts and temporary debug logs (removed during the rework).
Spec: requirement-swap-async.md

Arrangement-aware, non-mutating grouping (identityKeys -> targetKeys) that
generalizes useDataPoint.grouped. Foundation for moving grouping into the
worker. Unused for now.

The worker now owns the full raw dataset and treats arrangement, group,
sort, scale and showLabels all as compute params. After the one-time init
clone (and again only on a dataset switch), no setting change does any
main-thread data work — swap/group/sort/scale just post a param and render
the returned ChartData.

- transform.worker.ts: own raw rows; init/setArrangement project+group via
  projectAndGroup and reply with groupNames; compute reads grouped.get(group).
  Remove the broken swap message and swapAxisFields.
- useChartPipeline.ts: rawData/arrangement/labels/groupName inputs; reinit
  only on dataset change; arrangement posts setArrangement; group/sort/scale/
  labels recompute off the cache; expose groupNames. Drop triggerSwap.
- useDataPoint.ts: drop the 500k main-thread grouping computeds; add per-dataset
  arrangement state + activeArrangement; source groups from the worker via
  setGroupNames; keep resultGroups/activeGroupId/selectGroup working.
- Dashboard.vue: pass raw data + arrangement + derived labels + groupName;
  mirror the pipeline's groupNames into useDataPoint.
- AxisSwapper.vue: set the arrangement key instead of mutating rows/labels.
- swap.ts: drop the now-unused mutating swapAxisFields.

Groups are populated asynchronously by the worker on first `ready`.
Register a one-shot watcher (Vue 3.5 `{ once: true }`) on
`resultGroups.value.length` so the `?g=N` deep-link is applied once
groups arrive, without delaying `?d=` / settings which remain synchronous.

Replace groupName: Ref<string> param with activeGroupId: Ref<number> in
useChartPipeline. Resolve the group name internally via currentGroupName()
(reads from the pipeline's own groupNames ref), so pumpQueue() called inside
the ready handler uses the correct name synchronously — without waiting for a
downstream watcher to propagate it.

Previously, groupNames [] → [...] on first ready caused activeGroupName to
transition '' → groupNames[0] in a later microtask, which tripped the params
watch and triggered a full second recompute on every page load and every swap.
Now the params watch tracks activeGroupId (an integer), which does not change
on the first ready, so only one compute batch is queued. User group-select
(activeGroupId change) still triggers a single recompute as expected.

Also removes the now-unused activeGroupName computed from useDataPoint, and
collapses a double blank line in AxisSwapper.vue.

Switch dataSets to shallowRef+markRaw so rows stay plain objects;
postMessage clones them natively instead of the JSON round-trip
(~7300ms → ~700ms for 500k rows). Drop console.info(e) in the worker
that serialized the full message to devtools on every init.

Pre-populate chart slots in reinit() via listChartSignatures so per-card
skeletons appear immediately while the worker clones and groups data.
Remove the full-page LoadingSkeleton from the init phase — header and
metadata are visible from the moment data loads; only chart slots pulse.

Fix toStatSignature to handle absent per/unit fields without generating
misleading signature strings.