1.6.1

This is a maintenance release: nothing you see or use changes. The plugin,
the knowlery CLI, the MCP tools, the vault files and the 1.6.0 Jev ranking
behave exactly as in 1.6.0, and no new data is sent to Jev. 1.6.1 adds the
tooling that will measure whether a better cook-backlog ranking exists. It does
not ship that ranking: no new policy is enabled, and the ranking you get is
unchanged. The tooling lives in the repository only; it is not part of the
npm package or the Obsidian plugin.

Evaluation tooling (#86, #88)

• Design for measuring cook-backlog ranking. docs/design/compile-value/
  records the evaluation protocol (cv-eval@1, revision 3) and the signal-store
  design reviewed for 1.7.0. The 1.7.0 design is documentation only; none of
  it is implemented in 1.6.1.

• evals/compile-value/ harness. Four local scripts take a vault through
  one evaluation round:

  • freeze.ts copies the vault and verifies that the copy reproduces the
    vault's mtime order and its 1.6.0 Jev order;
  • sample.ts draws a stratified diagnostic sample and writes a blind
    annotation sheet;
  • pool.ts adds every available policy's top 20 to the sheet;
  • run.ts writes the metrics and the pre-registered decisions to a report.

  The scripts never call Jev, never write inside the repository, and refuse to
  overwrite sheets, private files or reports. In this round only the
  baselines (mtime order, the 1.6.0 Jev order, random, and an oracle ceiling)
  can be measured. The candidate policies need 1.7.0 producers, so every
  report states that no ship decision is possible.

• Fixes made while preparing the release. Actionable-critical estimates now
  count labelled pool notes exactly, as protocol §9.3 requires. Their intervals
  are withheld, not reported as zero width, when the sample has no event to
  resample. Reports record the harness version, so reports from before and
  after this fix can be told apart.

Data hygiene (#87)

• The Jev-triage experiment's real-vault labels were removed from the
  repository and replaced with a synthetic labels.example.yaml. The real
  file is now gitignored. The removed file remains in git history; rewriting
  history was deliberately not done.

Maintenance (#85)

• Every GitHub Action moved to its Node 24 major version. The release workflow
  itself is unchanged apart from the action versions.

Known limitations of the first evaluation round

• Intervals are exploratory for a single vault: they describe that vault's
  labelled notes, not other vaults.
• The pre-registered diagnostic veto compares age and type groups with the
  mtime baseline. Under mtime order, recent notes are at the top by
  construction, so any non-mtime candidate is likely to trip the age veto once
  the newest bucket has five or more sampled notes. Old notes are already near
  the bottom, so the veto barely protects them. The rule stays as registered
  for cv-eval@1; changing it requires a new protocol version before any
  candidate is judged.
• With about 80 sampled notes, actionable-critical notes outside the pool are
  rare. The bottom-20% non-inferiority checks will therefore often report
  insufficient evidence. This is reported as such, never rounded to a pass.