v0.1.0

Initial release: the full SPEC v3 implementation (Steps 1–24).

• Three-stage worthiness engine (hard rules → similarity classifier → LLM
  judge: ollama/deepseek/anthropic/openai, all optional) × scene engine with
  per-scene interrupt thresholds.
• Five routes: interrupt / digest / dispatch / curate / drop; dispatch
  verifies results before reporting ("done is a claim, not a proof") and
  arrives with a plan.
• Shadow mode (trust is earned), Tact Report, EMA learner with nightly
  human-readable POLICY.md distillation.
• Ingest protocol: webhook + MCP tools + GitHub/RSS pollers; delivery over
  terminal/desktop/Telegram with feedback buttons.
• Fully offline deterministic demo (uvx agent-chief demo) with a
  full-table routing regression.