CI regression harness: write→read round-trip + committed pixel-hash goldens

[cornish]

Why

Almost every bug this season lived in the write↔read seam — wsitools writes,
opentile-go reads, and the corruption only surfaced when something exercised the
full loop on a diverse input:

• wsitools#1 — convert --to {cog-wsi,svs,tiff,ome-tiff} dropped associated
  Predictor(317)/JPEGTables(347); decoded to garbage. Hidden because nothing
  read the emitted associated images back.
• VR OB-vs-OW (DICOM native PixelData) — emitted OW for 8-bit; only an
  independent reader exposed the RGB→grayscale collapse (the spike round-trip
  through the same lib missed it).
• The faithful-copy fix was only validated because I ran convert → re-read by
  hand, per format/codec.

These are exactly what a standing round-trip + golden harness catches on day one,
instead of by luck.

Proposal

1. Round-trip matrix (CI): for each fixture × writable target
   (svs|tiff|ome-tiff|cog-wsi|dicom), convert it, then re-read the output via
   opentile-go and assert the decoded pixels of L0 (+ each associated image)
   equal the source decode. This is the loop that exposed wsitools#1 / opentile-go#21
   / #23 — make it automatic. (Gate slow fixtures behind the integration tag /
   WSI_TOOLS_TESTDIR, like the existing integration suite.)
2. Committed pixel-hash goldens: hash --mode pixel already exists. Commit a
   per-fixture/per-format pixel-hash manifest and have CI assert stability, so a
   silent pixel regression fails loudly. (Pixel-hash, not file-SHA — the writer is
   nondeterministic at the byte level; see the deterministic-write issue.)
3. (Stretch) independent cross-reader check for the formats where opentile is
   both writer-validator and reader (TIFF family): re-decode emitted tiles with an
   independent decoder (tifffile / libjpeg / golang.org/x/image/tiff) — that's
   what definitively proved the VR-OB and predictor bugs were real vs reader
   artifacts.

Acceptance

• A convert of any corpus fixture to any writable target round-trips to
  pixel-identical associated images + L0 region (asserted in CI under the
  integration gate).
• Pixel-hash goldens committed; CI fails on drift.

Notes

• Pairs with the wsi-fixtures shopping list (wsi-fixtures#1) — more formats in the
  corpus = more of this matrix actually runs.
• The new tests/integration/convert_codec_test.go is a small step in this
  direction (novel-codec output re-read); this generalizes it to pixel-equality
  across all targets.

[cornish]

Depends on / pairs with the fixture corpus: WSILabs/wsi-fixtures#2 (master fixture wishlist). The more formats/encodings/dimensions that land there, the more of this round-trip + pixel-golden matrix actually runs in CI. The two are the systemic fix for the write↔read seam bugs (this repo's #1, opentile-go#21/#23/#7).

[cornish]

Extend to BIF once convert --to bif is round-trippable: a write→read pixel-golden for BIF. Caveat — opentile-go mis-renders BIF until opentile-go#57 lands (its serpentine read bug), so the BIF oracle should be bio-formats/QuPath (or a committed golden), not opentile. See the BIF parity set wsitools#11–#15.