Alluvia

Resilience release: provider rate limits no longer darken the map (#1, #2, #3).

• LLM governor — every provider call now runs behind a provider-agnostic
  governor: escalating cooldown ladder, per-model circuit breakers, and
  role→model fallthrough chains, all driven by abstract call outcomes (HTTP
  status and the standard Retry-After header are optional fast-paths, never
  dependencies). Breaker state persists in SQLite, so short CLI runs respect
  cooldowns learned by earlier runs instead of re-hammering an exhausted model.
• Fallthrough chains — on Groq's free tier each model has its own daily
  budget; distill/label/status/why calls fall through to sibling models when
  the head model hits a wall. propose never falls through — generation fails
  loud rather than silently downgrading. Override per role with
  ALLUVIA_LLM_CHAIN_<ROLE> (comma-separated).
• unfinished never goes dark — when the status classifier is
  unavailable, recurring themes get a heuristic status (recently touched →
  open, stale → dormant) instead of unknown. Heuristics are never cached;
  the classifier upgrades them on a later refresh.
• Degradation is visible — refresh prints a per-stage summary and a
  warning with the provider retry time when a stage degraded; themes /
  unfinished, the MCP tools, and the dashboard surface the same signal. An
  all-unknown map now says "the classifier hasn't completed" instead of
  pretending nothing is unfinished.
• Distillation pauses immediately (and resumably) when every model is cooling;
  fallback labels cut at word boundaries; adaptive pacing (AIMD) discovers a
  provider's sustainable request rate empirically.
• Breaking: every environment variable now uses the ALLUVIA_ prefix
  (ALLUVIA_DB, ALLUVIA_LLM_MODEL_<ROLE>, ALLUVIA_LLM_CHAIN_<ROLE>, …) — 0.1.0
  shipped them under an inconsistent prefix. If you exported variables for
  0.1.0, re-export them with the new prefix; config.toml users are
  unaffected.

Install: uv tool install alluvia · PyPI · Closes #1, #2, #3