v0.6.0

Fixed

• The shortlist tie-break was acting as a recency filter inside large equal-score bands. The
  ordered cut handed to the reranker keyed on score DESC, origin_count DESC, date DESC, and in a
  real pool the first two tiers tie across dozens of papers — so date decided, and the cap dropped
  central papers for being older than their band. On the recorded topic-2 acceptance run a paper
  screened 3 sat at rank 90 of a 289-strong in-window population and never reached the reranker; on
  the golden-topic run the cut fell inside the score-3 band and lost two. Latent since v0.4 and
  present on both eras of artefact — the evidence is in
  docs/measurements.md,
  552f09c462dce07a7c20fa3f30e85c3264f42346:research/experiments/phase12-selection/results/report_tail.md
  §4, and the audit record in
  #1. Fixed in
  1fd1465f413c21104d7af3710ed219ce595ca49a, with the replay that proves it in
  1fd1465f413c21104d7af3710ed219ce595ca49a:research/experiments/phase12-selection/results/src-t1-replay.json.
  Pooled golden survival into the rerank frontier at the shipped cap goes from 8/11 to 10/11,
  which is the finite maximum, with no cap change and no new weights.

Added

• A runtime deployment fingerprint: the server logs it at startup and /health reports it, so
  a running deployment can be matched to the commit it was built from. RELEASING.md gained the
  post-release step that exercises it — restart, then verify the fingerprint against the release
  SHA — and a release receipt.
• An experimental reference cognition driver in drivers/stateless/, repo-only. It screens a
  run's batches through stateless provider calls behind a reconciling CID contract, and every run
  writes a machine-readable engine provenance record. It is not part of the installed package:
  it lives outside src/, is excluded from the sdist, was never in the wheel, carries its own
  pinned dependency contract, and the package's own dependency list is unchanged. It informs the
  provider-neutral engine protocol that has not shipped
  (#4); promotion to a documented
  path requires
  end-to-end golden non-inferiority against a fresh multi-replicate conversational control, and
  cost is not a promotion criterion.

Changed

• Shortlist ordering is now a total order: score DESC, criteria_supported DESC, origin_count DESC, best_retrieval_rank ASC, date DESC, cid ASC. criteria_supported counts the
  distinct sub-criteria screening attributed to a paper (restricted to the ids queries.json
  defines, when the run has that file); best_retrieval_rank is the best position any source gave
  it. Both are evidence the pipeline already produced. The final cid tier removes the last
  dependence on candidates.json order — on the six frozen inputs it moves four rows, all inside
  fully tied bands, with membership at the shipped caps unchanged. No other selection semantics
  change: same caps, same threshold, same two windows, same slot rules at emit.
• The fields that order reads now have domains narrow enough to trust. score and
  Origin.rank are strict integers: Pydantic's lax coercion turned true into 1 and "3" into
  3, so a malformed screen.json bought a place in the order instead of exiting 2. A run whose
  artefacts carry a score or a rank as a string or a boolean now fails at the stage that reads it,
  with the field named. rank keeps its ≥ 0 floor — ranks are zero-based, and the top hit is
  rank 0. publication_date is unchanged at the schema, where partial dates are legitimate
  metadata, but the date tier resolves through a calendar check, so "2024-13-01" sorts with the
  unknowns instead of outranking "2025-01-01" as a raw string. And shortlist.build refuses a
  screen file that scores one cid twice, which both shipped call paths already refused upstream.