Street View: Fix depth map parsing when top left pixel is not sky - #45

…ad as gone

RED half of a test-first pair - all 7 new tests fail on the current seam,
and the flagship one fails INSIDE streetlevel's parser, demonstrating the
bug being routed around. They pin:

- The seam survives a payload whose first index byte is nonzero (a modelled
  zenith: tunnels, soffits, parking structures - the class streetlevel's
  uint16 offset misread kills on every attempt, unfixed through 0.12.11,
  fix pending in sk-zk/streetlevel#45), returning the correct raster in
  streetlevel order plus the plane bundle.
- An unrecognized envelope (an error JSON, a quota page that parsed) raises
  DepthPayloadError - transient, retried - and is never misread as "pano
  gone", which would ledger 'unavailable' and permanently write off a pano
  Google still serves.
- _compute_depth_raster: the mirror fixture reconstructed in payload order,
  all-no-plane when the plane list is empty, and PARITY with streetlevel's
  own decode on a payload upstream can read - so swapping the raster source
  changes nothing for the ~99% of panos both decoders handle.
- _decode_depth_planes rejects out-of-range plane indices (streetlevel
  fails the same payload with a bare IndexError mid-decode) and zero-area
  rasters.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

)

GREEN half of the test-first pair (7 red on the previous commit, 170 pass
now). Decided with Jon after his upstream fix (sk-zk/streetlevel#45) was
confirmed unmerged and absent from 0.12.11, the latest release and the
floor of our pin.

The seam no longer calls parse_panorama_id_response: its unguarded depth
decode misreads the payload's uint8 offset byte as a uint16, so any pano
with a modelled zenith (nonzero first index byte - tunnels, overpass
soffits, parking structures; a bit under 1% in a large sample) raised on
every attempt, was classed transient, and re-requested forever without
resolving. Now the seam reads the response code, the three orientation
scalars (with streetlevel's exact conversions), and the payload path off
the raw msg, decodes planes with _decode_depth_planes, and derives the
raster with _compute_depth_raster - the artifact's documented
reconstruction identity running forward, vectorized, with parity against
streetlevel's own decode pinned in CI for the payloads both can read. The
raster is handed over in streetlevel's mirrored order on purpose, so the
writer's #58 un-mirror and every CI pin on stored orientation are
untouched.

What remains of streetlevel in the depth path is api.find_panorama_by_id -
the protobuf-URL builder and fetch - with the session passed through
unchanged, so the request on the wire is identical. The WHY and the
revisit condition are written at the seam and in the offset canary test:
when upstream's fix ships and the pin moves past it, the bypass becomes
unnecessary rather than wrong - simplify back, or continue dropping the
dependency (#56 discussion).

Hardening that fell out: an unrecognized envelope (error JSON, quota page)
now raises DepthPayloadError - transient - instead of parsing as "pano
gone" and ledgering 'unavailable' for a pano Google still serves; the
decoder rejects out-of-range plane indices (streetlevel fails the same
payload with a bare IndexError mid-decode) and zero-area rasters. The
writer's mask invariant stays: no longer a cross-parser check, it now
guards the seam-to-writer flip plumbing and any raster/index divergence a
future edit introduces.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>