fix(wiki): recover from streamed chunk-repetition abort instead of failing the page (RES-947)

[jhkchan]

Problem (RES-947)

Dense OCR'd tables prime the wiki model to reproduce giant --- GFM separator walls. litellm's streaming repetition guard aborts with MidStreamFallbackError: ... The model is repeating the same chunk = ..., which propagated to the "subtopic page failed hard" branch → the page fell back to empty. This hits the exact content dominant in an architecture/engineering corpus (drawing schedules, dimension tables).

Live evidence (RLP POC): WikiCompiler: subtopic page failed hard (litellm.MidStreamFallbackError: litellm.InternalServerError: The model is repeating the same chunk = ------…).

Fix

Handle it at the source in _llm_generate_json. The repetition guard only runs in litellm's streaming handler, so when a streamed dispatch aborts with a repetition error, retry the call once non-streamed. The full content then flows to _call_llm's existing _is_degenerate_content dash-wall guard → deterministic Key Facts splice, rather than crashing. Fix lives in the shared helper, so every page kind inherits it (overview/people/activity/resources/glossary/subtopic).

• New _is_stream_repetition_abort(exc) — message-substring is the robust primary signal (MidStreamFallbackError isn't exported at litellm top level).
• New stream_override param on _llm_generate_json.
• Non-repetition errors re-raised unchanged; no retry when already non-streamed.

Tests

tests/wiki/test_compiler_stream_repetition.py: detector; streamed-abort → non-streamed retry returns good content (dispatch called stream=True then stream=False); non-repetition errors propagate; no retry when already non-streamed. Existing test_compiler_degenerate_guard + test_compiler_retry_gating still pass — 17 passed locally.

Part of epic RES-943 (RLP full-corpus scale + no-cloud gaps).

🤖 Generated with Claude Code

https://claude.ai/code/session_01PMMM6KQXmzAEA42UxpyMUm

[jhkchan]

Independent adversarial review — APPROVE (nits only).

No P0/P1 defect. Verified against the installed litellm: the real production exception from raise_on_model_repetition is litellm.InternalServerError with message "The model is repeating the same chunk = ...", and the detector's substring check fires on it. Detector does not over-match (ValueError, unrelated InternalServerError, CancelledError → False), the guarded litellm.exceptions import can't break module load, stream_override genuinely reaches dispatch_completion(stream=...), retry is exactly-once with correct re-raise and no infinite loop, no shared-state race.

Nits (follow-up):

• [P2] Docstring attributes the abort to MidStreamFallbackError, but that class is Router-only (litellm.Router) — since we call litellm.acompletion directly, that isinstance branch is effectively dead here; the real signal is the InternalServerError message substring. Reword; keep the isinstance as defensive-only.
• [P2 test fidelity] All retry tests raise a hand-authored RuntimeError("litellm.MidStreamFallbackError: ..."), not the real litellm.InternalServerError. Add one test using the real exception so a litellm message/rename can't silently break detection.