Region::fit does not discriminate at embedding dimensionality

[ops-ping]

A fitted Region either admits every probe or rejects every probe, at every rank and coverage tried. b_member and s_depth are shipped as first-class operators but do not appear usable at 768 dimensions with the exemplar counts any caller supplies.

Evidence

Proclamation sentences — 10-exemplar cloud of royal proclamations, probed with 4 held-out proclamations and 6 plain messages, through the real fit_region / b_member in vrules-wasm:

rank / coverage	proclamations inside	plain inside
3 / 0.90	4/4	6/6
3 / 0.95	4/4	6/6
5 / 0.95	4/4	6/6
8 / 0.95	0/4	0/6

Royalty terms — 24-exemplar cloud of royal terms, probed with 6 royal and 10 non-royal terms:

rank / coverage	royal inside	non-royal inside
3 / 0.90	6/6	10/10
5 / 0.95	5/6	10/10
8 / 0.95	5/6	10/10
12 / 0.90	5/6	10/10

A 9-exemplar version of the same cloud admitted tractor, coffee and bicycle while at rank 5 excluding royal woman and the queen of Spain — the wrong points on both sides.

Likely cause

depth_of_unit sums squared components along rank principal directions plus a residual term scaled by residual_scale. With a handful of exemplars in 768 dimensions the residual dominates and is near-identical for every vector (concentration of measure), so depth is nearly constant across probes and tau — a quantile of the cloud's own depths — lands inside that constant band. Raising rank toward the exemplar count inverts it: in-cloud residuals collapse, the per-direction scales shrink, and depths blow up for everything.

Impact

• Blocks using b_member for membership in the browser examples; a linear axis with c_project had to be used instead.
• crates/vrules-core/tests/fraud_triage.rs fits bec_phrasing_v1 from 6 exemplars at rank 3 and the Fraud triage example gates on it, so that region is likely non-discriminating too. Relevant to Re-enable Address, Fraud triage, Streaming and Proof once their runs assert behaviour #2.
• Existing tests pass because they only assert a region behaves consistently with the cloud it was fitted from, never that it excludes known negatives.

Suggested work

Establish empirically what exemplar count and dimensionality Region::fit needs to separate known negatives, and either document that floor or change the fit (e.g. shrinkage on the covariance, or dimensionality reduction before fitting). Add conformance coverage that probes a fitted region with negatives, not only with its own cloud.