docs: the RANSAC aligner was a defect, measured

AR-005 replaced cv::estimateAffinePartial2D(..., RANSAC, 3.0) with Umeyama
least squares over all five points — the estimator InsightFace aligns with, and
so the one the ArcFace/LVFace training crops were produced by.

The first note here assumed the two agree wherever RANSAC keeps all five points,
leaving a small divergence on non-frontal faces. Measured on 400 gallery
headshots with the model held fixed, that was wrong: the crops disagree by a
median 17 source px and 83.5% embed below cos 0.99 of their Umeyama counterpart.
A 4-DoF similarity is exactly determined by two points, so every minimal sample
fits its own pair perfectly and is scored on the other three; real landmarks sit
a median 2.74 canonical px from any similarity fit, so a landmark outside the
3 px band is the common case and RANSAC returns an under-determined transform.

How much that cost in accuracy is a separate question, and the honest answer is
less than those numbers suggest. Rebuilding the full gallery moved the
intra/inter separation the AR-023 calibration is fitted from by 0.583 to 0.590:
the old warp was wrong but self-consistent, gallery and probe both went through
it, and the embedder tolerates framing variation. The sharper evidence is
duplicate detection — the rebuild dropped 1614 near-duplicates against the
original build's ~100, because unstable two-point fits gave near-identical
images visibly different vectors. That instability, not a headline accuracy
delta, is what a tracker accumulating evidence across frames was paying for.

Also records the AR-030 residual's real-data floor: on the most cooperative
images the pipeline sees, it runs a median 2.74 px, so landmark noise occupies
the first few pixels and the synthetic foreshortening ladder is optimistic about
the low end. Any discount curve has to treat that range as uninformative rather
than as mild pose, and VR-012 must set thresholds against the measured
distribution.

Tests carry the tag they verify: the residual's roll/scale invariance and
monotonicity under foreshortening are what make it a pose measure rather than a
pose-and-everything-else measure.

TRACES: AR-005, AR-030 | SR-002

docs/SPEC.md

@@ -149,14 +149,55 @@ Produce the exact input ArcFace expects.
 **Current:** `align_face()` in `src/face_utils.hpp`, Umeyama fit via
 `umeyama_similarity()`, `cv::warpAffine` to `{112, 112}`. **Gap:** none.
 
-> **Migration note.** Until this landed the fit was
-> `cv::estimateAffinePartial2D(…, cv::RANSAC, 3.0)`. Where RANSAC kept all five
-> points its final refit is the same least-squares optimum, so the two agree;
-> they diverge exactly where a landmark fell outside the 3 px band — i.e. on the
-> non-frontal faces. Galleries baked before this change therefore carry embeddings
-> from a marginally different warp, concentrated on the hardest views. Rebuilding
-> is cheap and removes the question; GR-004's embedder stamp does **not** catch an
-> aligner change, only a model change.
+> **Migration note — this was a defect, not a refinement.** Until this landed the
+> fit was `cv::estimateAffinePartial2D(…, cv::RANSAC, 3.0)`. The expectation was
+> that the two agree wherever RANSAC keeps all five points, leaving a small
+> divergence on non-frontal faces. **Measured, that is wrong.** On 400 random
+> gallery headshots, one model held fixed and only the estimator varied:
+>
+> | | median | p90 | max |
+> |---|---|---|---|
+> | Crop disagreement (source px, over the crop corners) | 16.97 | 75.91 | 223.31 |
+> | `cos(umeyama, ransac)` for the resulting embedding | 0.791 | — | — |
+>
+> 83.5 % of crops embed to a cosine below 0.99 of their Umeyama counterpart —
+> they are not the same face crop. The mechanism is that a 4-DoF similarity is
+> exactly determined by **two** points, so every minimal RANSAC sample fits its
+> own pair perfectly and is then scored on the other three. Real landmarks sit a
+> median 2.74 canonical px from any similarity fit to the template (see AR-030
+> below), so images with a landmark outside the 3 px band are the common case,
+> not the exception; RANSAC then keeps two or three inliers and returns a wildly
+> under-determined transform.
+>
+> **Every gallery baked before this change must be rebuilt** — GR-004's embedder
+> stamp catches a model change, not an aligner change, so nothing else would say
+> so.
+>
+> **How much this cost in accuracy is a separate question, and the answer appears
+> to be: less than the crop numbers suggest.** Rebuilding the full gallery
+> (2456 actors) moved the intra/inter separation the AR-023 calibration is fitted
+> from only slightly:
+>
+> | | intra-actor | inter-actor | separation |
+> |---|---|---|---|
+> | RANSAC | 0.6234 | 0.0407 | 0.5827 |
+> | Umeyama | 0.6340 | 0.0440 | 0.5900 |
+>
+> The reconciliation is that the old warp was *wrong but self-consistent*: it
+> produced a differently-framed face rather than a scrambled one, gallery and
+> probe went through the same estimator, and the embedder tolerates framing
+> variation. So the figures in `model-bakeoff.md`, `best-model.md` and
+> `pose-expansion.md` were all produced through the broken warp on both sides and
+> should be re-run, but there is no measured basis for expecting them to move far.
+>
+> The sharper evidence of the old instability is duplicate detection: rebuilding
+> with an unchanged `dedup_tol` dropped **1614** near-duplicate images, where the
+> original build dropped on the order of a hundred. Near-identical source images
+> used to embed to visibly different vectors — RANSAC fitting two-point subsets is
+> unstable under small landmark perturbations, and being RNG-driven it was not
+> reproducible either. That instability is what a tracker accumulating evidence
+> across frames pays for, and it is the strongest reason the fix is worth having
+> independently of any accuracy delta.
 
 ## AR-006 — Embedding
 
@@ -237,6 +278,15 @@ the same response.
 Only size drops the face outright, and only because VR-005 measured a knee below
 which the embedding carries no signal to discount. Blur and pose are different:
 
+  **The synthetic ladder is noise-free and therefore optimistic about the low
+  end.** Measured on 400 real TMDB/Jellyfin headshots — the most frontal, most
+  cooperative population the pipeline ever sees — the residual runs p5 1.11,
+  median 2.74, p90 4.82, max 6.35 canonical px. So landmark noise alone occupies
+  roughly the first 3 px, and the synthetic sweep's "26° yaw ≈ 1.2 px" sits
+  *below* the noise floor on real data. VR-012 must set any threshold against
+  this measured distribution, and a discount curve has to treat the first few
+  pixels as uninformative rather than as mild pose.
+
 - A blurred or turned face is still evidence of **presence**, which is what
   SR-002 actually asks about.
 - The tracker admits a link on position *or* identity precisely so that a face