A quote and its time words are found across a line break in the passage

[WaylandYang]

Why

Found in an end-to-end run on dev. An English board minute, as a PDF, lost every date: "will step down as Chief Technology Officer on December 31, 2024" became a statement with no time.

The PDF's text layer breaks lines at the margin, so the passage reads on December⏎31, 2024. The model copies the sentence back with a space. locate compared byte for byte, so the quote was not found in the chunk and the date was not found in the quote. The drops table said so: 5 × quote_not_in_chunk, 3 × time_not_in_quote for one page.

What changes

• locate lets a run of whitespace match a run of whitespace. Everything else is compared as before, and the offsets it returns are still character positions in the passage.
• locate_time finds the time words in the passage's own sentence rather than adding the position inside the copy, since the two can differ in how long the whitespace is (a break plus an indent against one space).

A word broken in the middle (Septem⏎ber) still does not match: that is a different word as far as the check can tell.

How it was checked

• New unit test: a sentence and its date found across a break with an indent, the offsets pointing at the passage; the reverse case; a mid-word break and a changed date still refused.
• End to end on a real server with deepseek-flash: a PDF whose lines break between words. Before: no dates and 8 drops. After: 2024-12-31, 2025-01-01 and 2024-09-01 on their statements and no drops.
• cargo test -p utopia-server extraction_open, cargo clippy -p utopia-server --all-targets -- -D warnings, cargo fmt --all --check.

Before review

• Every commit is signed off
• fmt, clippy and the affected tests pass

🤖 Generated with Claude Code