[asaf-shitrit]

Why

Closes #1095. Open extraction kept an OutOfCredit refusal as a string and moved to the next chunk, so the job used all 3 attempts and no llm.out_of_credit alert fired.

What changes

call_and_parse_one_chunk returns the error with its chain intact when utopia_llm::out_of_credit finds it, next to the no-chat-model case. unextracted stays a list of strings, and nothing outside the helper changes.

Afterwards, chunks applied before the refusal stay in the graph with extracted_at set. The document is failed, with the refusal as its graph_error. The next run continues from the refused chunk. Like the other errors inside the chunk loop, this exit skips the end-of-run refresh_disambiguators.

How it was checked

A test in extraction_open_out_of_credit_tests.rs uses a scripted endpoint and a real database. It extracts 3 chunks, and the endpoint refuses the second one. The third chunk is not called, the error is hopeless at attempt 1 of 3, and one llm.out_of_credit alert fires. The fact from chunk 0 stays, only chunk 0 has extracted_at, and the document is failed.

The test fails on dev without the fix.

Before review

• Every commit is signed off (git commit -s)
• cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings and cargo test --workspace pass

[WaylandYang]

Thank you. This is the small fix we agreed on #1095: one arm in call_and_parse_one_chunk returns the refusal with its chain intact, and nothing outside the helper moves.

The test covers what the issue asked for: the third chunk is not called, the error is final at the first attempt, one llm.out_of_credit alert fires, and the chunk applied before the refusal stays. Thank you also for stating what the document looks like afterwards, including that this exit skips the end-of-run disambiguator refresh like the other errors in the loop.

Merging after the release sync that is in flight. Thanks again.