v2.2.0 — descriptions that say what they are not for, and a way to measure it

Fifteen merges since v2.1.0, fourteen of them pull requests. Nothing here changes what a
plugin does at the surface. What changed is how reliably the right one fires — and whether
that claim can be checked rather than asserted.

Routing

• Every description now says what its skill is not for (#9, #22). Six were rewritten
  after a blind router measured them misfiring; the rest gained a boundary clause. That is
  why the average description grew from 173 to 204 characters — the trade that bought fewer
  misfires.
• The description's fourth copy is now checked (#11). The same string is catalogue copy in
  /plugin, marketplace.json, the plugin README.md blockquote and the skill's own
  frontmatter; the blockquote is line-wrapped, so an edit by exact string match used to reach
  three of them and leave the fourth stale.

Measurement

• make eval drives claude plugin eval over 13 committed cases (#20) — the misses
  from a 3-sample blind-router baseline that scored 367/380. It stays outside make test
  on purpose: each case is a full Claude child session on your own credential, so it wants a
  network and about $5.50, and CI here has no secret to give it.
• All 19 trigger-eval sets hardened or relabelled (#18, #21, #23, #27). Two hardened
  negatives were themselves caught by the router run and resolved in opposite directions —
  one a mislabel, one a description finding that was withdrawn rather than kept as a query
  engineered to pass.
• What the two runners actually measure is now recorded (#26). The blind router is a
  forced choice among 19 descriptions; make eval observes whether the Skill tool was
  invoked at all. So a blind-router miss is a hypothesis and a make eval miss is an
  observation — and a skill should not fire when the assistant can simply comply in the
  current turn.

Skill bodies

• Every phase states what it does when its precondition fails (#19), as one shared
  paragraph per skill rather than 76 separate stop conditions.
• references/ is in use for the first time (#12), by agent-smith's Phase 5 — Autonomous
  mode only, so a Partner-mode dispatch never pays for its scoring table.
• agent-smith Phase 2 and 2.5 have stated conditions instead of (if needed), with a
  tie-break for when both hold.

Release mechanics

• The version-bump rule is settled and written down (#17): a bump tracks what a plugin
  does. Catalogue copy alone moves the top-level version, and a plugin moves at most once
  per release window.

Seventeen plugins move to 2.1.0. briefing and shortcut stay at 2.0.0, having no
phases to change.

Full changelog: v2.1.0...v2.2.0