[LLM] Add canonicalization/auto-fix stage for model output

[Simonsbs]

Goal

Constrain model output to canonical L0 more aggressively.

Scope

• Add canonicalization and structural auto-fix pass between model output and verifier.
• Strip markdown/noise robustly.
• Normalize section order, ids, and instruction formatting when mechanically safe.
• Reject non-repairable output with explicit error taxonomy.

Acceptance

• Canonicalization pass has dedicated tests.
• Verify success improves measurably without lowering semantic correctness.
• Failure classes are reported in benchmark output.

[Simonsbs]

Implemented core #7 work in this round:\n- added generated-output sanitization stage (fence removal, CRLF cleanup, module extraction from first section)\n- added canonicalization/auto-fix stage using between generation attempts\n- added per-task and aggregate to benchmark reports\n\nNext: add stricter structural auto-fixes and explicit non-repairable failure taxonomy refinement.

[Simonsbs]

Correction/update for this issue:

Implemented core #7 work in this round:

• Added generated-output sanitization stage (fence removal, CRLF cleanup, module extraction from first ver section).
• Added canonicalization/auto-fix stage using l0c canon between generation attempts.
• Added per-task last_error_class and aggregate error_class_counts in benchmark reports.

Next:

• Add stricter structural auto-fixes and clearer non-repairable taxonomy labels.

[Simonsbs]

Update: LLM benchmarking docs/wiki now support both Ollama and OpenAI adapter workflows. Model leaderboard and trends docs were refreshed; qwen2.5:3b currently leads local tested set while qwen3:8b timed out repeatedly under the current local setup.

[Simonsbs]

OpenAI matrix snapshot published (commit a5adb29). Models tested: gpt-4.1-mini, gpt-4.1 with L0 adapter. Results: verify 33.3%, semantic 16.7%, avg attempts 1. Recommended model in current snapshot: gpt-4.1-mini. Wiki/docs updated and published.