feat: banded admission for the per-subject embedding store

AR-018 — an embedding joins a track's store only if its similarity to something
already there falls inside a band, rather than merely being far from the gallery.

Above the upper bound it is redundant: another look at a pose the store already
covers, teaching the annex nothing while costing a slot a novel view could have
used. Below the lower bound it is suspect: within one track every face is the
same person by construction, so an embedding unlike everything else on the track
is evidence that construction failed — a track-ID collision or a bad detection.
Admitting it is exactly how an actor's annex gets poisoned with someone else's
face.

The old gate had only the upper half of that idea, expressed as a raw cosine
against the gallery. Both bounds are now calibrated probabilities (AR-024), so
the same number means the same thing here as in association and evidence
weighting rather than three different things.

This catches track-ID collisions EARLIER than the spread gate did — at the door
rather than at promotion — so the buffer never becomes two-person in the first
place. The spread gate stays as a second line for a track that drifts gradually
instead of jumping. The existing test was asserting the mechanism rather than
the outcome, so it was rewritten to assert what actually matters: whichever gate
fires, the outsider must not reach the annex.

Rejections are counted. A store that admits nothing is as broken as one that
admits everything, and neither is visible otherwise.

Band defaults 0.90-0.95 are working values pending VR-007; the two bounds fail in
opposite directions and must be swept separately.

Suite: 92 cases, 6133 assertions.

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

TRACES: AR-018, AR-024 | SR-005

docs/SPEC.md

@@ -133,9 +133,30 @@ Produce the exact input ArcFace expects.
   nose, left mouth, right mouth).
 - Alignment is the *only* geometric normalisation; no additional augmentation at
   inference.
+- **The transform is fitted by Umeyama's closed-form least squares over all five
+  points**, which is what InsightFace uses (skimage's `SimilarityTransform` *is*
+  `_umeyama`) and therefore what produced the crops ArcFace and LVFace were
+  trained on. The canonical warp is part of the input distribution, not an
+  implementation detail (AR-011).
+- **Not a robust estimator.** A RANSAC fit buys a small residual by discarding
+  the landmarks that disagree with the model, and on a turned face those are the
+  foreshortened ones — the signal AR-030 reads. With five points and a two-point
+  minimal sample it also cannot separate a mis-detected landmark from honest
+  out-of-plane rotation, so the robustness is nominal while the cost to AR-030 is
+  total. It is RNG-driven besides, which made replay determinism a property of
+  thread scheduling.
 
-**Current:** `align_face()` in `src/face_utils.hpp:9-22`, `cv::warpAffine` to
-`{112, 112}`. **Gap:** none.
+**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.
 
 ## AR-006 — Embedding
 
@@ -174,15 +195,43 @@ the same response.
   the crop is already scale-normalised, so a measure taken there cannot silently
   re-measure face size and double-count it against AR-002.
 - **Visibility** — extreme pose or occlusion means the face presents fewer of the
-  features the embedding assumes are present. Derived from the **5-point
-  landmarks AR-001 already emits** — nose offset from the eye midpoint over
-  inter-ocular distance, plus eye/mouth-corner asymmetry — which are already
-  computed, already used by AR-005, and already in the VR-001 dump, so the
-  measure costs one arithmetic expression per face and can be studied on existing
-  fixtures with no GPU. A dedicated landmark model (`models/2d106det.onnx` is
-  present but referenced nowhere) is **not** adopted unless VR-012 shows the
-  5-point proxy insufficient: an extra inference per detection is precisely the
-  cost AR-011 says not to spend.
+  features the embedding assumes are present. The measure is the **residual of
+  the AR-005 alignment fit**: the RMS landmark error, in canonical 112×112
+  pixels, left over after the best similarity transform onto the ArcFace
+  template. It costs nothing — the transform is computed for the warp regardless,
+  and the residual is what that fit could not explain.
+
+  Two properties earn it the job over an explicit yaw estimate:
+
+    - A similarity absorbs rotation, uniform scale and translation **exactly**,
+      so the residual is by construction the non-similarity part of the
+      deformation: out-of-plane rotation and foreshortening. In-plane roll
+      contributes nothing, so "a tilted head reads as a turned one" is excluded
+      structurally rather than by tuning. The destination frame is fixed, so face
+      size cannot leak in either — that is AR-002's axis, and double-counting it
+      would make a small frontal face look occluded.
+    - It responds to **occlusion** and to plainly broken landmark sets, which an
+      angle regressor by construction does not: a hand across the face is not a
+      rotation, but it does displace landmarks.
+
+  Indicative magnitudes from a synthetic foreshortening sweep (`k ≈ cos yaw`):
+  `k=1.0 → 0.00`, `0.9 → 1.18`, `0.75 → 3.11`, `0.5 → 6.72`, `0.3 → 9.85`
+  canonical px. Smooth and monotone with a usable range; the mapping onto real
+  faces is VR-012's to establish, and no threshold is set from these numbers.
+
+  Neither a dedicated landmark model (`models/2d106det.onnx` is present but
+  referenced nowhere — and it emits points, not pose) nor a direct pose CNN is
+  adopted unless VR-012 shows the residual insufficient. If one is needed the
+  candidate is **6DRepNet** (MIT, RepVGG-B1g2, 3.47° MAE on AFLW2000) rather than
+  Hopenet, which it dominates on accuracy, licence, recency and export
+  friendliness. Two caveats to record before that happens: both are trained on
+  **300W-LP**, which inherits research-only terms from 300W's constituent sets,
+  and both want their own loosely-framed ROI rather than the ArcFace crop — a
+  second warp and a second image in flight, which lands on AR-004's byte-based
+  backpressure gap. It would also have to run **per track** — over the bounded
+  view set AR-019's diversity buffer already keeps — not per face per frame,
+  which is the cost rule applied as written: fewer regions, never a degraded
+  input.
 
 **Failing an axis discounts the observation; it does not delete the detection.**
 Only size drops the face outright, and only because VR-005 measured a knee below
@@ -216,14 +265,18 @@ hand-chosen cutoff on an uncalibrated measure is the same unfalsifiable magic
 number AR-024 retired for similarity, and it would fail the same way: meaning
 something different for every detector, every embedder and every film.
 
-**Current:** none of the three is assessed. `min_face_px` (40, decoded-frame
-space) is the only quality signal in the pipeline; sharpness and visibility are
-unmeasured, and `align_face()` silently drops only the degenerate-affine case
-without counting it.
+**Current:** visibility is measured and carried — `estimate_alignment()` in
+`src/face_utils.hpp` returns the residual alongside the transform, and
+`FaceAlignerFunc` writes it to `DetectedFace::alignment_residual`. Size is
+`min_face_px` (40, decoded-frame space — AR-002 still open). Sharpness is
+unmeasured. Nothing yet *consumes* any of it: no discount is applied, and
+`align_face()` still drops the degenerate-fit case without counting it.
 
-**Gap:** all of AR-028 … AR-030. Order: land the quality vector and its dump
-field first (AR-028) so VR-012 can be run from fixtures, then set behaviour per
-axis from what it measures.
+**Gap:** AR-029 entirely. For AR-030, the measure exists but the discount does
+not — it must reach `EvidenceDiscounter` as the reliability term. For AR-028, the
+residual does not yet reach the VR-001 dump, which is what VR-012 needs to run
+from fixtures; that is the next step, since it unblocks the study that sets
+every remaining behaviour.
 
 ## AR-007, AR-008 — Tracking