Demosaic a Fujifilm sensor instead of refusing it

Every RAF stopped at the embedded preview, because the demosaicer had
one kernel and it was a Bayer kernel. D11 makes Fujifilm first-class and
FR-RAW-5 asks for it by name, so a hard error there was a promise we had
not kept.

X-Trans is a 6x6 tile, and nothing in the Bayer path survives that: the
missing channels sit at different offsets at all 36 positions, so there
is no fixed kernel to write. The new shader fits a weighted plane
through each channel's samples in a 5x5 window and carries the other two
channels across as the difference between those planes, keeping the
pixel's own measured value untouched. A plane rather than a mean because
the three channels are sampled at different places in the tile: a mean
compares a red taken slightly left of the pixel with a green taken
slightly right of it, and that offset is a colour cast that follows
every gradient in the frame. The fit is done in white-balanced space,
where the constant-colour-difference model it rests on is actually true
of a neutral subject; that alone halves the error at a luminance edge.

Two compromises, both deliberate.

It is not Markesteijn. There are no directional hypotheses and no
homogeneity map, so it does not resolve detail finer than the CFA period
and a hard edge arrives about two pixels wide. It cannot ring — the
output is bounded by the local sample range — so it does not produce the
worms FR-RAW-5 exists to avoid, but the quality that requirement asks
for is still owed.

The tile's phase is guessed rather than known. rawler has each body's
pattern exactly, as a 36-character string, but CfaPattern::XTrans throws
it away before dr-gpu sees the file, and it is not a constant to
hard-code: the bodies in that database start the tile at four different
origins. So the phase is read back out of the pixels, by grouping the 36
per-position means and taking the grouping with the least spread. That
part needs nothing from the scene. Telling red from blue does — shifting
the tile by half a tile turns it into itself with red and blue swapped,
so no geometry can decide it — and the as-shot white balance is what
breaks the tie. A frame that is almost entirely one colour can defeat
that; widening dr-decode to carry the pattern string would retire the
guess altogether.

The tests assert reconstruction, not success: a flat patch comes back
exactly at all six phases tested, and a linear ramp comes back exactly
too, which is the property the plane fit exists for and the one a mean
would fail.

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

core/dr-gpu/src/error.rs

@@ -22,9 +22,10 @@ pub enum GpuError {
     #[error("readback failed: {0}")]
     Readback(String),
 
-    /// The X-Trans demosaic is not implemented (FR-RAW-5). Reported rather
-    /// than approximated with the Bayer path, which would produce a maze of
-    /// colour artefacts and look like a corrupt file.
+    /// The CFA layout is one no demosaic here handles — in practice a file
+    /// `dr-decode` could not identify the pattern of. Reported rather than
+    /// approximated with the Bayer path, which would produce a maze of colour
+    /// artefacts and look like a corrupt file.
     #[error("unsupported CFA pattern: {0}")]
     UnsupportedCfa(String),