Add the A4 report designer and fix band layout

Brings in the A4 (millimetre) report designer alongside the existing
matrix editor, migrates the frontend from react-scripts to Vite/Vitest,
and fixes four defects found while using the A4 band tools.

Layout — layout.js now implements the standard banded-report model:
everything prints where it was drawn, and a band growing per data row is
the only thing that moves anything. Previously `paginate` re-flowed every
band from the top margin, so a band placed halfway down the page
previewed at the top, and `header` was treated as a page header and
hoisted above the statics drawn before it. A header is a COLUMN header:
it prints in place and repeats only on a continuation page. Statics now
take part too — one drawn below a band keeps its drawn gap as that band
grows, and follows onto the next page rather than vanishing when pushed
past the bottom margin. The displacement is keyed on the band's design
bottom, not on an instance's laid-out y: every row of a band shares one
design position, so keying on the instance measured the gap from row 1
and let the static overlap rows 2..n.

Band children — a child's x is relative to its band, as its y already
was. layout() emitted it raw, so a field printed measured from the paper
edge instead of the band's left margin. Instances now carry bandX.

Band placement — a band's x comes from the page margins, so creation
takes only the click's Y (the factory's x was being overwritten by the
spread order) and dragging is locked to the vertical axis. Bands render
as a strip, with top/bottom rules only, matching the matrix editor.
fromApiReport reflows on load so reports saved with a stray x straighten
themselves. Matrix bands are constrained to vertical too.

Drag lifecycle — useElementDrag re-subscribed its document listeners on
every render during a drag, because the effect depended on an unmemoized
onDrag. Each re-subscribe removed the mouseup listener before adding the
next, measured as add/remove/add/remove/add across two mousemoves, and a
release landing in that gap was lost: the element kept following the
cursor until the next click. Snapping made it routine, since a held
guide stops x/y changing and leaves the guide overlay's render sitting
between the move and the release. onDrag and dragData now go through
refs so the effect depends on [isDragging] alone.

Note: the matrix editor has no test coverage, so its band constraint is
verified by reading only. The A4 suite is 201 passing tests.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

.claude/agent-memory/react-expert-dev/a4-smart-guides.md

@@ -0,0 +1,35 @@
+# A4 smart guides (`a4/guides.js`)
+
+Pure functions, no React/DOM. `buildGuideCandidates(elements, excludeIds, pageSetup)`
+→ `{V, H}`; `resolveSnap(draftRect, candidates, zoom)` → `{x, y, guides}`;
+plus `unionRect(boxes)` and `snapToleranceMm(zoom)`.
+
+## The threshold is in SCREEN PIXELS, not mm
+
+`SNAP_PX = 6`, converted per-call via `SNAP_PX / (PX_PER_MM * zoom)`. This is
+load-bearing, not a detail: a fixed mm threshold feels sticky zoomed in and
+useless zoomed out. Do not "simplify" it to a constant mm value. The property
+worth asserting is `tolMm * PX_PER_MM * zoom === SNAP_PX` at every zoom.
+
+## Integration seams in `A4Canvas.jsx`
+
+- Candidates are built **once** in `handleElementDragStart` into a `useRef`.
+  Rebuilding per mousemove is the one thing that makes dragging laggy.
+- Multi-select snaps the **union rect**, then applies one delta to every member.
+  Snapping members independently tears the group apart.
+- `hooks/useElementDrag.jsx` is shared with the matrix editor and forwards only
+  pixel deltas — **never widen it**. Alt-to-suppress is tracked with a
+  window-level `keydown`/`keyup` listener in A4Canvas, plus a `blur` reset (an
+  alt-tab otherwise leaves snapping permanently off, because the keyup lands in
+  another window).
+- Bands emit no candidates (their edges *are* the margins, already candidates),
+  but band children do.
+
+## Gotcha: snapping breaks drag tests that assume raw deltas
+
+Any pre-existing test that drags an element and asserts the exact px→mm result
+can fail once a candidate happens to sit within ~1.6mm of where it lands. This
+bit `a4Elements.test.js`'s rect case: it drifted 0.05mm from the 20mm left
+margin and legitimately snapped. The fix is to hold Alt around the drag
+(`withAltHeld` helper in that file), not to weaken the assertion — the test is
+about unit conversion, so switching the magnet off states its intent.