v0.5.0

0.5.0 - 2026-07-31

soon 0.5.0 makes the measured contextual policy the local default and adds an
optional model-backed candidate layer without putting inference or network
access on the ordinary prediction path. Suggestions remain editable text and
are never executed automatically.

Added

• Added a contextual full-command ranker with separate candidate-source
  and ranker boundaries, smoothed transition and context evidence, repository
  and branch capture, feedback-aware ranking, explainable debug output, and a
  chronological promotion comparison against the v0.4 baseline. It is the v0.5
  source default after passing the quality and latency gate; v0.4-baseline
  remains an explicit fallback (#16).
• Added opt-in local and OpenAI-compatible candidate sources for bounded
  rerank, Repair, and explicit Generate modes. Provider context is minimal and
  filtered, model output is treated as untrusted, deadlines fall back to the
  deterministic policy, and Zsh/replay retain source and outcome attribution
  (#17).

Full diff