1ru.7: Make LM Studio and Ollama first-class local endpoint integrations

[penso]

Automated handoff for 1ru.7.

Issue:
Base branch: main
Head branch: beads-polyphony-1ru.7
Commit: dba0627

[penso]

Summary

The branch makes ollama and lmstudio usable as first-class openai_chat providers, with sensible default local base URLs and no required API key. The main review concern is around edge-case behavior rather than the happy path.

Risks

• crates/agent-openai/src/lib.rs:118, crates/agent-openai/src/lib.rs:288, crates/agent-openai/src/lib.rs:362: the new "skip unreachable local provider" path depends on reqwest timeouts or connection-refused errors, but OpenAiRuntime still builds a default client with no request timeout. If a local endpoint accepts the connection and then stalls, model discovery can still hang instead of degrading cleanly.
• crates/workflow/src/render.rs:347: LM Studio auth fallback only checks LMSTUDIO_API_KEY. LM Studio documents LM_API_TOKEN for bearer auth, so authenticated LM Studio setups will not work out of the box unless users duplicate the token into repo-specific config or set api_key explicitly.
• crates/workflow/src/render.rs:472, crates/agent-openai/src/lib.rs:103: OLLAMA_BASE_URL, LMSTUDIO_BASE_URL, and explicit base_url overrides are accepted verbatim. There is no validation that the override still points at the expected OpenAI-compatible /v1 base, so a typo like http://host:1234 will fail later with a fairly opaque runtime error.

Recommended human checks

• Smoke-test both providers against real local servers, including lmstudio with fetch_models = true.
• Try a dead-but-listening local endpoint once, not just a refused port, and confirm discovery behavior is acceptable.
• Verify an authenticated LM Studio setup using only LM_API_TOKEN, or document the required env var alias before merge.