Add output contract adaptation for v0.2.2

[HuRuilizhen]

Summary

This PR is the second v0.2.2 slice for strengthening the LLM concept-extraction path.

The main goal is to make output shaping an explicit part of the provider contract so that llm extraction can adapt across stronger and weaker response-format capabilities without changing the downstream ConceptMention contract.

Included

• adds explicit LLM output-contract modes

  • introduces auto, json_schema, json_object, and prompt_only
  • includes the selected output-contract mode in cache keys
  • keeps auto as the default behavior

• adapts provider requests across contract modes

  • tries json_schema first in auto mode when a schema is available
  • falls back to json_object, then to prompt-only requests
  • also falls back when stronger contract modes return empty provider content

• hardens extractor parsing across weaker contract modes

  • accepts fenced JSON blocks
  • can recover the last valid {"paragraphs": ...} object from free-form text
  • preserves the same downstream paragraph-aligned concept parsing contract

• adds focused contract-mode coverage

  • provider tests cover explicit json_object and prompt_only modes
  • provider tests cover weaker-contract fallback after structured-output rejection
  • provider tests cover weaker-contract fallback after empty content
  • extractor tests cover fenced JSON and free-form reasoning text recovery

Validation

Repository checks:

• ruff check .
• pyright
• pytest -q

Observed local-model check:

• validated the contract fallback path against a local Ollama service
• the current qwen3.5:4b model still does not reliably emit a usable JSON answer for the bundled example prompt, so this PR improves the compatibility layer but does not claim full local-model prompt stability yet

Notes

• this PR intentionally focuses on output-contract adaptation and parser resilience; broader benchmark work remains the next 0.2.2 slice
• local-only planning documents under docs/, along with local experiment material, remain intentionally untracked and are not part of this branch

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 75e54fc875

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".