feat(session): retry transient provider failures

[pepicrft]

Summary

• Wraps each LLM call in the native runtime with a bounded retry policy (default: 3 attempts, exponential backoff + jitter on HTTP 408/425/429/500/502/503/504 and stream errors wrapping those).
• Streaming calls only retry while no event has been emitted to the caller, so partial output is never silently duplicated. Once a chunk has shipped, the underlying error is surfaced and the caller decides how to recover.
• Configurable via the :retry option on Condukt.run/3, Condukt.stream/3, and start_link/1 (boolean, keyword list, or %Condukt.Retry{} struct). Also accepts a custom :classify callback.

Why

In production we just hit a 503 service overloaded from an upstream LLM in the middle of a streaming Slack reply. The session bubbled the error straight to the user as a generic fallback message even though the call was a textbook transient failure. Retrying once at the provider boundary would have masked it entirely; in the steady state every Condukt consumer wants the same behavior.

Test plan

• mix test — 405 passed, 37 excluded
• mix credo lib/condukt/retry.ex lib/condukt/session.ex lib/condukt.ex — only the pre-existing 32-field struct warning
• mix format --check-formatted clean for changed files
• Wire it through Atlas (separate PR) and confirm an injected 503 retries and recovers

🤖 Generated with Claude Code