[codex] Add OpenAI Responses fallback for mini models

[nyldn]

Summary

• add OpenAI Responses API routing for GPT-5.x / Codex model names when users explicitly select them with --backend openai --model ... or GRAPHIFY_OPENAI_MODEL
• add opt-in OpenAI model fallback via GRAPHIFY_OPENAI_FALLBACK_MODELS, preserving the existing gpt-4.1-mini default and Chat Completions path
• document mini-model examples such as gpt-5.5-mini with fallback to gpt-5.4-mini,gpt-4.1-mini

Scope / duplication check

• This does not change backend detection priority or the existing OpenAI default model.
• I checked open upstream issues and PRs for openai codex mini fallback gpt-5.5 gpt-5.4; no matching open item came back.

Validation

• uv run --with pytest pytest tests/test_llm_backends.py tests/test_ollama.py -q -> 18 passed
• python3 -m compileall graphify/llm.py graphify/__main__.py -> passed
• uv run graphify --help | sed -n '/extract <path>/,+8p' -> verified help text
• graphify update . -> completed code graph refresh
• uv run --with pytest pytest -q -> 667 passed, 6 failed in existing Fortran/SQL parser tests unrelated to this backend change (test_fortran_capital_F_parses_preprocessed, SQL table/view/function/edge assertions)

[safishamsi]

Two issues to fix before merging:

1. _call_llm not updated for Responses API routing
extract_files_direct routes GPT-5/codex-mini to _call_openai_responses, but _call_llm (used by the dedup tiebreaker path) still routes everything through _call_openai_compat. GPT-5 users hitting the dedup path will get API errors. _call_llm needs the same if model in RESPONSES_API_MODELS branch.

2. Replace assert last_exc is not None with an explicit raise

# instead of:
assert last_exc is not None
raise last_exc

# use:
if last_exc is None:
    raise RuntimeError("retry loop exited without exception")
raise last_exc

assert is silently removed under python -O, which would cause an UnboundLocalError on last_exc in optimized builds.

Fix these two and CI passes → ready to merge.