[doc] update docs

doc/tl-annotation-stats-v1.4.0.md

@@ -55,6 +55,70 @@ threshold. Above 21 px recall is ≥0.96 and above 101 px it is ≥0.996 — the
 the fit guards are removing sub-resolution speckle in the masks, not lesions. Read the
 corpus number as "0.547 of components", never as "0.547 of findable lesions".
 
+### Recall climbs with physical size — ≥20 mm disease is ≥0.97 everywhere
+
+`tl_recall_by_length.csv` re-stratifies the same exactly-attributed clusters by physical
+extent instead of pixel count. The per-cluster length is the **max Feret diameter of its
+pixel centres in millimetres**, each in-plane axis scaled by its own spacing (43.8% of
+slices are anisotropic in-plane, so a pixel-space diameter would mis-size them). It comes
+from the same mask recount as the denominator, with spacing read from the NIfTI header
+zooms, so it exists for every cluster — unmeasured ones included, which the fitted major
+axis by definition cannot cover. Its margins reproduce the corpus numbers exactly
+(6,951,667 clusters, 3,801,540 measured, 0.547).
+
+| Feret length | Clusters | Measured | Recall |
+| --- | ---: | ---: | ---: |
+| ≤2 mm | 2,232,427 | 1,952 | 0.001 |
+| 2–5 mm | 974,881 | 409,412 | 0.420 |
+| 5–10 mm | 955,724 | 694,683 | 0.727 |
+| 10–20 mm | 1,055,235 | 976,105 | 0.925 |
+| 20–50 mm | 1,263,241 | 1,250,268 | 0.990 |
+| 50–100 mm | 413,816 | 412,849 | 0.998 |
+| >100 mm | 56,343 | 56,271 | 0.999 |
+
+This is the clinical-units counterpart of the pixel table above, and the direct check of
+what the v1.4.0 millimetre floor set out to do. **Recall is 0.967 at ≥10 mm and 0.992 at
+≥20 mm, and at ≥20 mm no dataset is below 0.97** — the release
+[blog post](https://medvision-vlm.github.io/blog/tl-annotations-v140.html)'s headline
+claim ("above 97% in every one of the 12 collections, and 99% overall"), reproduced here
+per dataset:
+
+| Dataset | ≥20 mm clusters | Measured | Recall |
+| --- | ---: | ---: | ---: |
+| KiPA22 | 11,297 | 11,297 | 1.000 |
+| LIDC-IDRI | 8,039 | 8,036 | 1.000 |
+| HNTSMRG24 | 36,959 | 36,936 | 0.999 |
+| MSWAL | 73,392 | 73,341 | 0.999 |
+| MAMA-MIA | 155,306 | 154,901 | 0.997 |
+| BraTS24 | 705,471 | 702,719 | 0.996 |
+| DEEP-PSMA | 84,086 | 83,675 | 0.995 |
+| LNQ2023 | 17,718 | 17,542 | 0.990 |
+| KiTS23 | 66,580 | 65,835 | 0.989 |
+| MSD | 307,790 | 303,468 | 0.986 |
+| autoPET-III | 254,062 | 249,289 | 0.981 |
+| PI-CAI | 12,700 | 12,349 | 0.972 |
+| **All 12** | **1,733,400** | **1,719,388** | **0.992** |
+
+The ≥20 mm ranking has almost nothing to do with the overall-recall ranking: MSD, worst
+overall at 0.336, reads 0.986 here, and DEEP-PSMA goes 0.645 → 0.995. Datasets differ in
+their speckle load, not in how well sizeable lesions are measured. Of the 3,150,127
+unmeasured clusters, 88.8% are ≤5 mm, 97.0% are ≤10 mm, and 14,012 (0.44%) are ≥20 mm.
+
+The ramp between 2 and 20 mm is the resolution-adjusted floor, not a leak: v1.4.0
+measures a cluster when its *fitted* major axis clears max(2 mm, 2 × the coarser in-plane
+spacing), so the 2 mm term binds only on sub-millimetre grids — a 3 mm-spacing reformat
+keeps nothing under 6 mm — and physical recall climbs over a decade instead of stepping
+at 2 mm. Two properties of the binning itself:
+
+* The bin variable is the Feret proxy, not the fitted axis, and the two straddle the
+  floor differently (pixel centres understate a contour's extent by about one pixel),
+  which is why the ≤2 mm bin is not empty by construction: 1,952 measured clusters land
+  there.
+* States are paired with lengths positionally through the pixel-count multiset, and two
+  same-count clusters on one slice are interchangeable under it — counts and recall stay
+  exact, but which of the two lengths goes with "measured" is arbitrary at that
+  degeneracy. `ambiguous = 0` in every bin, as in the pixel table.
+
 ### The axial/reformatted gap tracks anisotropy
 
 On thick-slice datasets the axial plane is near-complete while the reconstructed planes