fix: add retry-with-feedback loop to oo learn for invalid TOML

[randomm]

Task Description

oo learn frequently generates invalid TOML. Research shows the most reliable cross-provider fix is a validation + retry loop: if the LLM returns invalid TOML, send the parse error back to the LLM and ask it to fix only the broken part.

Root cause: LLMs occasionally generate almost-valid TOML but fail at specific lines (trailing commentary, malformed sections). This is a known problem with all LLM-based structured output generation.

Research findings (2026-03-05):

• Anthropic structured outputs (JSON schema constraint) is 100% reliable but JSON-only — requires JSON→TOML conversion
• Retry with error feedback works across ALL providers (Anthropic, OpenAI, Cerebras)
• Temperature 0 reduces variability
• Stop sequences prevent trailing commentary

Proposed Solution

In src/learn.rs, in run_learn(), after calling the LLM:

1. Validate the TOML response
2. If invalid, send a follow-up message: "Your previous TOML was invalid: {error}. Here is what you returned: {toml}. Output ONLY the corrected TOML, nothing else."
3. Retry up to 2 times
4. If all retries fail, write to learn-status.log and return Err

Also add temperature: 0 to API calls for consistency.

Quality Gates (Non-Negotiable)

• TDD: Write tests before implementation
• Coverage: 80%+ test coverage for new code
• Linting: All code passes project linting rules
• Local Verification: oo learn cargo test produces a valid pattern file

Acceptance Criteria

• run_learn retries up to 2 times on invalid TOML with error feedback
• Temperature set to 0 on all LLM API calls
• oo learn cargo test reliably produces a valid cargo-test.toml
• Tests cover the retry logic (with mock LLM responses)
• No changes to provider routing — works for Anthropic, OpenAI, Cerebras