Record the new Best in the spec and the manual

denoise.md §15: why Best became one network, how it compares with the
mixture and Medium on real photographs and the chart, the candidates that
fell short, the file names, the saved-edit numbering, and the timings on
the 3050 -- 0.51-0.54 s whole-frame, 0.95 s in tiles, against 2.60 s for
the mixture in tiles. The manual lists Bilinear, Fast and Best, says an
edit made with Medium opens with Best, and gives the new time; Medium's
close-up goes.

docs/dev/denoise.md

@@ -571,3 +571,59 @@ In PyTorch fp16 the same network over the whole 20 MP frame in one call took 3.5
 tiles, so a card that holds a whole frame gains more than the 6 GB one does; `WHOLE_FRAME_MAX` is a
 constant sized for 6 GB until the limit follows the card's memory.
 
+## 15. Best becomes one network (0.24)
+
+The photographer's goal for 0.24 was Best's quality in under a second on the laptop. Whole frames
+(§14) took the mixture from 2.60 s to 1.37 s and no further on a 6 GB card, so the other half was
+a single network that holds the mixture's quality at a third of its work. Methods are now
+`Bilinear`, `Fast` and `Best`; Medium and the mixture are retired.
+
+**The network** is `fb-combo` (darkroom-denoise, 2026-10-07): §11's shape (32-64-128-192, blocks
+1-1-2-2, 3.2 M parameters, 48 GMAC/MP, halo 192), 20 000 steps from `fb-edges2` ← `student-m`,
+taught by the mixture at a half share, with 10 % drawn scenes and 25 % crops from the edge-rich
+cells of the training frames (branch `edge-sampling`). Scored on real photographs — the chart
+overstated the mixture's lead (a chart-sharp network was softer than Medium on real edges) — on the
+validation crops in the top quarter for sharp detail:
+
+| | Edge PSNR, ISO 1600 / 6400 / 25600 | Sharpness kept | Smooth areas | Held-out PSNR, ISO 400 / 1600 / 6400 / 25600 | Chart edge |
+|---|---|---|---|---|---|
+| Mixture (Best to 0.23) | 30.71 / 29.93 / 28.55 | 0.899 / 0.868 / 0.782 | 42.61 / 41.71 / 40.12 | 40.69 / 39.84 / 38.47 / 36.71 | 0.82 |
+| `fb-combo` (Best from 0.24) | 30.67 / 29.87 / 28.49 | 0.902 / 0.874 / 0.792 | 42.54 / 41.57 / 39.85 | 40.64 / 39.78 / 38.38 / 36.54 | 0.89 |
+| Medium (to 0.23) | 30.43 / 29.68 / 28.38 | 0.896 / 0.862 / 0.776 | 42.59 / 41.69 / 40.08 | 40.59 / 39.75 / 38.40 / 36.65 | 1.30 |
+
+Edges within 0.04–0.06 dB and more sharpness kept at every ISO; the known shortfall is smooth areas
+at ISO 25600, 0.27 dB. The photographer took it as it stood at 20 000 of a planned 30 000 steps.
+Others tried on the way, each short of the mixture on real photographs: `fb-sharp` (drawn scenes,
+chart-sharp but Medium's real edges), `fb-edges` (half edge-rich crops: edges close, ISO 25600
+flats −0.24 dB), `fb-edges2` (a quarter: 0.03–0.14 dB short everywhere, chart 1.33–1.47), and a
+from-scratch 24-48-96-128 between Fast and Medium.
+
+**Files.** `mosaic-hq-1408.onnx`, `mosaic-hq.onnx` (any size) and `mosaic-hq-1408.a16w16.onnx` for
+the Hexagon. A new name, not Medium's or Best's: the result cache keys a model by name and size,
+and this one is byte for byte Medium's size. The tablet form lost 0.00 dB in simulated QDQ at every
+ISO and at most 0.09 dB across the ×0.5–×4 noise bracket (A16W8 0.08 / 0.26 dB; int8 −10.6 dB);
+not yet confirmed on the tablet itself.
+
+**Saved edits** keep their numbers: 2, which was Medium, is now Best; 3, which was Best, is past the
+end and reads as the default, Best. Both land on the new network with no migration.
+
+**Measured** 2026-10-07, `_MG_8862`, RTX 3050 Laptop, TensorRT fp16, P3 / 5001 MHz, nothing else on
+the card:
+
+| Best | Network time | Peak GPU memory |
+|---|---|---|
+| mixture, 1408² tiles (0.23) | 2.60 s | — |
+| mixture, whole frame (§14) | 1.37 s | — |
+| `fb-combo`, 1408² tiles | 0.95 s | 0.55 GB |
+| `fb-combo`, whole frame (two 4160 × 3248) | **0.51–0.54 s** | 1.75 GB |
+
+Decode and the hot-pixel pass add 0.4–0.5 s, so a photograph is about a second end to end. Whole
+frame against tiles: max |Δ| 0.0029, 90 dB apart — fp16's spread. The whole-frame engine's first
+build took 12 minutes (the mixture's 28); from the cache it loads in about a second, so the session
+keeps the engine's ordinary 30 s idle decay rather than unloading after each photograph: at 1.75 GB
+it fits beside the develop view on a 6 GB card, and an unload would cost the next photograph a
+second.
+
+The manual's close-up for Best is still the mixture's render, which this network matches to within
+the table above; it is re-recorded with the next pass of `tools/manual/record.sh`.
+

docs/manual/README.md

@@ -215,22 +215,22 @@ colour that come with it, while keeping the fine detail. Look at it at
 
 `Method` chooses how:
 
-- `Best`, the default: two networks, one for smooth areas and one for
-  edges, blended where each is better. The cleanest skies and the sharpest
-  lettering, and the slowest.
-- `Medium`: one network taught by `Best`. Nearly as clean in smooth areas,
-  a little softer on hard edges, in about a third of the time.
-- `Fast`: a smaller one, taught the same way. Visibly noisier at very high
-  ISO than the other two, but still far cleaner than none, and quick.
+- `Best`, the default: clean skies and sharp lettering, edges kept as
+  crisp as the camera recorded them.
+- `Fast`: a smaller network, taught the same way. Visibly noisier at very
+  high ISO than `Best`, but still far cleaner than none, and quicker.
 - `Bilinear`: the camera's ordinary conversion, noise and all.
 
+A photograph last edited with `Medium`, which earlier versions offered,
+opens with `Best`.
+
 The photograph shows the camera's ordinary conversion while the network
 works, with its progress in the bar at the top, and changes when it is
-done — on a laptop's graphics card, about two and a half seconds for a
-20-megapixel photograph with `Best` and under one with the other two;
-longer on a processor alone or on the tablet. The first photograph after
-installing waits a few minutes more while the graphics card prepares each
-network, once. The result is kept, so a photograph opened again,
+done — on a laptop's graphics card, about a second for a 20-megapixel
+photograph with `Best`, reading the file included; longer on a processor
+alone or on the tablet. After installing, the graphics card spends up to a
+quarter of an hour preparing each network, once, in the background; the
+photographs developed meanwhile take a little longer. The result is kept, so a photograph opened again,
 or exported, does not wait a second time, and switching back to a method
 already used is quick.
 `Strength` eases it off: below 100 % it puts back some of what was removed,
@@ -241,8 +241,8 @@ The lamp and railing of a night frame at ISO 8000, at 1:1, by each method:
 | Bilinear | Fast |
 |---|---|
 | ![The railing and the lamp at ISO 8000, as the camera recorded them](media/develop-denoise-bilinear.png) | ![The same, with the Fast network](media/develop-denoise-fast.png) |
-| **Medium** | **Best** |
-| ![The same, with the Medium network](media/develop-denoise-medium.png) | ![The same, with the Best network](media/develop-denoise-best.png) |
+| **Best** | |
+| ![The same, with the Best network](media/develop-denoise-best.png) | |
 
 It works on raw files from any camera with the usual colour pattern of
 red, green and blue squares — not on JPEGs, and not yet on Fujifilm's

docs/manual/index.html

@@ -296,22 +296,21 @@ colour that come with it, while keeping the fine detail. Look at it at
 1:1, where noise lives.</p>
 <p><code>Method</code> chooses how:</p>
 <ul>
-<li><code>Best</code>, the default: two networks, one for smooth areas and one for
-edges, blended where each is better. The cleanest skies and the sharpest
-lettering, and the slowest.</li>
-<li><code>Medium</code>: one network taught by <code>Best</code>. Nearly as clean in smooth areas,
-a little softer on hard edges, in about a third of the time.</li>
-<li><code>Fast</code>: a smaller one, taught the same way. Visibly noisier at very high
-ISO than the other two, but still far cleaner than none, and quick.</li>
+<li><code>Best</code>, the default: clean skies and sharp lettering, edges kept as
+crisp as the camera recorded them.</li>
+<li><code>Fast</code>: a smaller network, taught the same way. Visibly noisier at very
+high ISO than <code>Best</code>, but still far cleaner than none, and quicker.</li>
 <li><code>Bilinear</code>: the camera's ordinary conversion, noise and all.</li>
 </ul>
+<p>A photograph last edited with <code>Medium</code>, which earlier versions offered,
+opens with <code>Best</code>.</p>
 <p>The photograph shows the camera's ordinary conversion while the network
 works, with its progress in the bar at the top, and changes when it is
-done — on a laptop's graphics card, about two and a half seconds for a
-20-megapixel photograph with <code>Best</code> and under one with the other two;
-longer on a processor alone or on the tablet. The first photograph after
-installing waits a few minutes more while the graphics card prepares each
-network, once. The result is kept, so a photograph opened again,
+done — on a laptop's graphics card, about a second for a 20-megapixel
+photograph with <code>Best</code>, reading the file included; longer on a processor
+alone or on the tablet. After installing, the graphics card spends up to a
+quarter of an hour preparing each network, once, in the background; the
+photographs developed meanwhile take a little longer. The result is kept, so a photograph opened again,
 or exported, does not wait a second time, and switching back to a method
 already used is quick.
 <code>Strength</code> eases it off: below 100 % it puts back some of what was removed,
@@ -319,8 +318,8 @@ as grain without colour, for a picture that does not look too smooth.</p>
 <p>The lamp and railing of a night frame at ISO 8000, at 1:1, by each method:</p>
 <table><thead><tr><th>Bilinear</th><th>Fast</th></tr></thead><tbody>
 <tr><td><img src="media/develop-denoise-bilinear.png" alt="The railing and the lamp at ISO 8000, as the camera recorded them" /></td><td><img src="media/develop-denoise-fast.png" alt="The same, with the Fast network" /></td></tr>
-<tr><td><strong>Medium</strong></td><td><strong>Best</strong></td></tr>
-<tr><td><img src="media/develop-denoise-medium.png" alt="The same, with the Medium network" /></td><td><img src="media/develop-denoise-best.png" alt="The same, with the Best network" /></td></tr>
+<tr><td><strong>Best</strong></td><td></td></tr>
+<tr><td><img src="media/develop-denoise-best.png" alt="The same, with the Best network" /></td><td></td></tr>
 </tbody></table>
 <p>It works on raw files from any camera with the usual colour pattern of
 red, green and blue squares — not on JPEGs, and not yet on Fujifilm's