fix(street-level): KartaView PR review — context loss mid-load, chunked reads, source switches, a11y

From the IC8 review of #1121 (2 P2, 7 P3; each with a test that fails on
the old code):

P2
- A WebGL context lost while a newer panorama loaded left a black viewer
  and a 'successful' open: the panorama now builds on the new canvas (or
  falls back to its front view).
- The streaming body read every real answer takes was tested with one
  chunk only: answers split mid-character are reassembled; the shared test
  fake now streams, so every source test runs the production path.

P3
- Lighting another source by its chip (params) now stops a running lookup
  and clears its spinner and error, as setProviderEnabled did; a restore
  that changes nothing leaves them alone.
- Panoramas decode off the main thread (ImageBitmap, RGBA upload: ~90 ms →
  ~15 ms at 4,096 px on a phone-class CPU) and are 2,048 px docked, 4,096
  once the viewer grows; a closed viewer keeps a 1×1 texture.
- The lit KARTAVIEW chip label is #8f8aff (~6:1; #635bff was 3.7:1).
- The viewer's keyboard focus ring is drawn above the photo.
- Tests: a second context loss and a drag on the new canvas; the look
  direction kept across a loss; a page's 30 s first try.
- The browser gate loses a real WebGL context in the 360° step and checks
  the 429 lookup was asked again (the fixture counter is reset first).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

scripts/fixtures/street-level/kartaView.mjs

@@ -33,6 +33,11 @@ export function resetKartaViewFixture() {
   nearestLookups = 0;
 }
 
+/** Nearest lookups answered since the last reset (the first got a 429). */
+export function kartaViewLookups() {
+  return nearestLookups;
+}
+
 /** The drive's photos, south to north; every tenth is 360°, every odd one archived. */
 export function kartaViewPhotos() {
   const step = KV_SPACING_M / METRES_PER_DEG_LAT;