v0.13.10

Added

• LLM provider seam + claude-cli subscription backend (#330). A new
  athenaeum.provider module centralizes LLM client construction behind
  build_llm_client(config) and resolve_provider(config) (env
  ATHENAEUM_LLM_PROVIDER > yaml llm.provider > api). Two first-party
  backends:
  • api (default): wraps anthropic.Anthropic(...) verbatim — params pass
    through unchanged, so prompt caching (#230), the Messages Batch API
    (#236), retries, and cost accounting are byte-for-byte identical to before.
  • claude-cli: drives the operator's ambient Claude Code subscription
    login via claude -p --system-prompt <sys> --model <id> --output-format json. No credential handling (same ambient-auth stance as the git-push
    path, #284). The adapter mirrors client.messages.create(**params) so the
    compile-path call sites (tiers, contradictions, resolutions) are
    unchanged; the recall-time query_topics preprocessor stays on the api
    path by design (a per-recall subprocess would add seconds to every query).
    Constraints: cache_control is stripped on the CLI path (preserved on
    api); CLI rate-limit / timeout / transient exits map to
    _retry.TransientAPIError — caught downstream as a give-up so the affected
    file is deferred to the next run (not retried in-run); a missing claude
    binary fails loudly at startup; batch mode is API-only and claude-cli +
    ATHENAEUM_BATCH_MODE is a loud startup error (no silent fallback); and
    claude-cli token COUNTS are still recorded in TokenUsage while
    estimated_cost_usd reports $0 (subscription-covered). See
    docs/configuration.md → "LLM provider selection (#330)".