Five fixes from manual QA:
1. ObservationCard — clicking anywhere on the card now selects it.
The wrapping button only covered the text rows; empty space was dead.
The whole card is now role="button" with keyboard (Enter/Space) support;
inner action buttons (↑ ↓ 🗑) stop propagation so they don't double-fire.
This was also the root cause of "details look stuck" — the active
observation simply wasn't being switched.
2. NextAppointmentPicker — when value === undefined the "Date" radio was
still showing as checked because `kind` defaulted to 'date'. The three
radios now read from the raw `value?.kind` (which can be undefined),
so "Aucun" is the only checked option when no next RDV is set.
3. svgToPng.rasterizeSvg — `canvas.toBlob` can return null in some
browsers when the canvas state is unusual; the call now falls back to
`canvas.toDataURL('image/png')` and decodes the base64 payload. Also
paints a white background before canvg renders (so the PDF embed is
opaque) and clamps width/height to >= 1px to avoid zero-area canvases.
canvg errors are caught + logged instead of propagating, so the export
still produces a (possibly blank) page rather than throwing.
4. svgConvention.parseSvgBones — clickable bones are now distinguished
from decorative ones. Interactive = has a hitbox element OR category
is spine-zone OR head. Decorative bones (teeth, malar, mandible,
nasal-bone, facial-crest, individual vertebrae paths…) get
`pointer-events: none` applied to their elements and are excluded from
the returned map. Result: in Horse_head.svg, frontal/temporal/parietal/
sphenoid/occipital/atm/ssb/c0-c1/hyoid become clickable as intended,
while teeth / malar / mandible no longer intercept the clicks.
5. (covered by #1) — the per-observation details panel always read the
active observation; the bug was the click handler scope. Confirmed by
switching observations: the editor section now updates immediately.
88 tests still pass, lint clean, build clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>