v0.38.0-beta — a second dense retriever, published

A second dense retriever is now PUBLISHED, not described

0.37.0-beta named the dense retriever and disclosed, in prose, that the per-shape flag moves under
a different embedding model. The consuming project replied that they would derive that second column
themselves from our compare tool. Two derivations of one number can drift, and a drift between a
consumer's copy and ours is invisible to both, so it is published instead.

No corpus content changed. All ten corpus_sha256, every question id and every question count
are identical to 0.36.0-beta and 0.37.0-beta. 587 questions, 36 shapes, 10 verticals.

Additive — nothing existing moved

Verified field by field against the previous release: 210 pre-existing values compared, 0
altered. allgold_dense, predicted_headroom_dense and discriminates_under_dense keep their
exact meaning and value. They remain the reference column.

New in probes.retriever_sensitivity:

field
second_dense_retriever	azure-emb-text-embedding-3-small-d1536-cosine-f16
second_dense_note	what the pair is for, and that neither retriever is the reader's
by_shape.<shape>.second_dense	allgold, predicted_headroom, discriminates
by_shape.<shape>.retriever_agreement	robust-ranking | retriever-sensitive | non-ranking

Every shape now carries a verdict, including the one that cannot be measured

forgetting/never-known had no retriever_agreement. All 15 of its questions have an empty gold
set — the correct answer is an abstention — so gold.issubset(top_k) is vacuously true and ALLgold
would read 1.000 under every retriever: the most flattering number available and the least true.
It is now declared rather than omitted:

"never-known": {
  "questions": 15,
  "retrieval_measured": false,
  "retriever_agreement": "not-applicable",
  "not_measured_because": "..."
}

No allgold or headroom fields are written for it, deliberately — a row that cannot be averaged by
accident. Sidecars now cover all 36 shapes: 18 robust-ranking, 10 retriever-sensitive, 7
non-ranking, 1 not-applicable.

The three-class read, across 35 shapes

class	shapes	meaning
robust-ranking	18	discriminates under both published retrievers
retriever-sensitive	10	the two disagree on discrimination or on whether dense beats BM25
non-ranking	7	neither discriminates

The 10 sensitive shapes are 4 flag-flips (forgetting/still-valid,
temporal/interval-position, workingmemory/distance-25, workingmemory/distance-40) and 6
sign-flips (arithmetic/delta, episodic/list-order, prospective/due-window,
temporal/recency, workingmemory/distance-60, workingmemory/distance-8), and the two sets are
disjoint — checked, not assumed.

Two guards on the publication itself

• The reference column is recomputed and compared to what is already published. Same quantity,
  two independent computations, never previously put side by side. A mismatch on any of the 70
  comparisons stops the publication rather than overwriting the older number. All 70 agreed.
• --stamp runs the plumbing self-check for each model first and refuses on anything but an
  exact 0.000000 self-spread. A compare tool is a probe, and a probe that cannot come out the
  other way publishes noise.

Costs zero API calls: both models' vectors were already banked, and this only re-ranks them.