v0.14.0

Added

• keel review <slug> and per-criterion human judgement. G3 used to let one
  merge approval stand in for every human oracle. Each human oracle is now
  discharged per criterion and per run, against a brief of what the reviewer
  saw, kept as run evidence. (ADR-0028)
• Oracle sensitivity. G2 checks that an oracle fails without the change it
  claims to verify, so assert!(true) no longer passes as evidence. The check
  is native and blocking; sensitivity_paths in keel.toml says where it
  looks. Mutation testing through cargo-mutants is a plugin and advisory.
  (ADR-0027)
• What the agent spent. Drivers report the tokens they actually used, and
  keel metrics shows them beside what keel injected. The two are labelled
  and never summed: one is reported, the other is estimated. (ADR-0026)
• docs/comparison.md: a plain account of where keel fits, and where
  another tool is the better choice.
• ADR-0026 to ADR-0028 record the three decisions above.

Fixed

• A run that predates the current approvals no longer satisfies the Run
  stage. After a spec or plan was re-approved, the old passing run still
  counted, so keel next jumped to the merge approval for a change that had
  not been rebuilt. keel next now says the run is stale and names it, and
  keel next --json carries "stale_run" (additive; the keel.next/1 schema
  string is unchanged).