Run each model on the Hexagon in the form measured to hold it

The engine knew f32 and int8, and gave the Hexagon int8 for every role it
served. Measured on the tablet itself (inference.md §1.5), int8 lost
5% of the detector's faces at 40-80 px, moved the landmarks 1.5 px,
emptied the segmenter's scores and cost the denoiser 5-9 dB; fp16 the HTP
refuses outright. `Form` gains A16W8 and A16W16, and `Rung::form` now
names one per role: detectors and landmarks A16W8, the segmenter, scene
model, border filler and denoiser A16W16, XFeat int8. The embedder and
the eye classifiers stay on the CPU.

Each loader resolves its `<stem>.<form>.onnx` sibling; the segmenter and
XFeat, compiled into the binary, embed their quantised forms on Android
only and pick through `choose_embedded`. The probe, the compile step and
the cache fingerprint follow the form instead of assuming int8. Detectors
on the new form write `scrfd_*_a16+w600k_mbf`, and `model_ids` answers
for all three spellings.

On the tablet (ORT 1.29 + QNN 2.42), each shipped file against f32 on the
same inputs, and against the CPU's f32 time:
  SCRFD 500m/2.5g/10g  A16W8   100% of faces in every band   4.2/5.1/9.0 ms vs 17/56/198
  landmarks            A16W8   0.25 px in the 192 crop        0.5 ms vs 2.8
  YOLO26n-seg          A16W16  98.2% found, mask IoU 0.994    12.9 ms vs 90
  scene model          A16W16  98.9% of cells agree           15 ms vs 151
  MI-GAN               A16W16  41 dB from f32 in the fill     87 ms vs 488
  XFeat                int8    pano alignment 0.45 px (f32's own spread 0.41)  6.5 ms vs 58
  denoiser             A16W16  0.00 dB at every ISO            95 ms vs 1510 a tile
Face numbers are over public COCO val2017 photographs, not a library.

The APK carries the siblings (BUNDLED 15 -> 19; the old int8 detectors
removed), about 43 MB more. The Windows installer and its CI count skip
them; the Arch and Flatpak packages list their files and never had them.
The ladder example takes a role per model, which is how the per-role
forms above were seen landing on the NPU from the real probe.

docs/dev/panorama.md

@@ -467,9 +467,11 @@ the tablet. Three ways to make it viable, none built:
 1. **Fill at a quarter of the resolution and upsample.** Sky and scree
    tolerate it; twenty-odd tiles, about three minutes on the desktop CPU. A
    background job with the outbox's patience, not an interactive one.
-2. **int8 on the tablet's Hexagon through QNN**, where the plain-conv design
-   is the point and the whole graph should run in milliseconds. The setup
-   exists from the eye-state work; MI-GAN is a candidate for the same path.
+2. **The tablet's Hexagon through QNN**, where the plain-conv design is the
+   point. Measured 2026-10-04 (inference.md §1.5): int8 changes the fill
+   (16 dB from f32's), so it ships with 16-bit activations and weights —
+   87 ms a tile against 488 ms on the tablet's CPU, the whole graph on the
+   NPU, 41 dB from f32 in the hole.
 3. **A WGSL runtime for those six operators.** A project of its own, and
    the only route that would make it interactive on the desktop.