fix(dng): repair the real-camera conformance tier and gate it per-PR

[justin13888]

Closes #353.

Both feature halves of #353 are already merged — #410 (typed ColorProfileInfo/NoiseProfile
read projection) and #411 (gamut-dng consumes the gamut-metadata facade; ExifMetadata
removed). This is the real-camera conformance tier catching up to them, plus the CI gap that let
it fall behind.

The breakage

#411 removed DecodedDng::exif_extra — correctly, since the whole EXIF sub-IFD now arrives inside
DngMetadata::exif and an extras list beside it would be the same data twice. But
tooling/gamut-dng-real-conformance/examples/probe.rs still read that field:

error[E0609]: no field `exif_extra` on type `DecodedDng`
   --> examples/probe.rs:104:17

Extended CI's "Real camera DNG conformance" job was green on the #410 merge and has failed on
every push since #411. Because the failure is at compile time, the gate has never run a
single sample against #411's change — the corpus assertions, the Adobe DNG SDK differential and
the preserving-rewrite round-trip have all been dark since that merge.

Why no per-PR gate caught it

gamut-dng-real-conformance is excluded from the workspace and nothing depends on it — it is
invoked by manifest path from mise run test-dng-real. So cargo clippy --workspace --all-targets and cargo test --workspace never compile it, and both #410 and #411 validated
clean. It is the only excluded tooling crate with this property: every other one is a path
dev-dependency of a workspace crate, so the workspace build reaches it.

ci: adds mise run check-dng-real to the lint job to close that. cargo check needs no corpus,
so the job does not grow the ~178 MiB fetch, and the expensive half — the Adobe DNG SDK — is
already built there for gamut-dng's own dev-dependency on gamut-dng-oracle.

The probe change

Beyond the repair, probe now reports the two projections #353 added. probe is how a corpus file's
MANIFEST.toml expectations are derived from measured behaviour, and these tags previously
reached it as the ifd0_extra/raw_extra residue it already prints — so typing them made probe
quieter without making it more informative. Measured output:

  profile       : model="iPhone13,3 back camera" as_shot_neutral=[0.448…, 1.0, 0.540…]
  color_profile : hue_sat=[-, -, -] look=- tone_curve=257 pts exposure_offset=None
      third calibration set: matrix=false illuminant=None calibration=false forward=false
      reduction matrices: 1=- 2=- 3=-
  noise_profile : planes=1 [scale=3.000e-5 offset=3.000e-8]
  extras        : ifd0=8 raw=1 exif=31 gain_map=true gain_map2=false

exif=31 is the EXIF sub-IFD's own entry count, not an extras list — the field the counter now
reads is the facade's whole directory, where the pre-#411 ExifMetadata carried five.

The Adobe-converted Canon file exercises the table path (hue_sat=[90x30x1 Linear, 90x30x1 Linear, -] look=36x8x16 Linear), and the Leica M Monochrom — no colour calibration at all —
prints color_profile : none / noise_profile : none, which is the case #410 called out for
keeping noise_profile off CameraProfile.

Validation

Command	Outcome
mise run check-dng-real	pass — the E0609 is gone
mise run test-dng-real	6 passed, 0 failed (408s), incl. the Adobe SDK differential
mise run lint	pass
mise run fmt-check	pass
mise run check-commits	no errors in 2 commits
probe on 4 corpus files	new lines render, populated and empty cases

test-dng-real was run against the full fetched corpus — the first time it has executed against
#411's change.

Notes

• Deliberately touches nothing under crates/: no Cargo.toml, CHANGELOG.md or STATUS.md,
  which chore: release #412 (the open chore: release) owns. Commit types are test/ci so release-plz derives
  no version bump while that PR is in flight; none of the three paths sits in a published package.
• Pre-existing and left alone: cargo fmt --all does not reach this excluded crate either, so
  gaps.rs, src/lib.rs, tests/real_corpus.rs and parts of probe.rs carry nightly-rustfmt
  deviations. Same root cause, but reformatting them here would bury a 3-file fix in unrelated
  churn. Worth a follow-up that decides whether the tier should be fmt-gated too.