fix(knowledge): tolerate whitespace in verbatim citation-quote check

[vanbrady]

What changed and why

The cited-summary step compared model quotes against chunk text with an exact substring match (quote not in chunk.text). PDF extraction keeps the source's hard-wrap newlines inside a chunk, so a word pair split across a printed line stays broken by a newline, while a faithful model quote renders that break as an ordinary space. Every such quote failed the check, so the semantic PaperFlow(PaperSemanticCfg(...)).build(...) raised a citation error for any real paper regardless of model — the same failure the PaperFlow docstring recorded for the 2026-07-24 run.

This adds quote_matches_chunk_text in the knowledge layer as the single rule for "the quote is present in its chunk": it compares under collapsed whitespace, so a quote that differs only by whitespace runs counts as verbatim while quotes whose words or order differ are still rejected. All three checks now route through it:

• flows/_paper_summary._validate_research_draft (research-draft validation)
• PaperGlobalSummary.from_draft (citation resolution)
• PaperSemanticResult reload validator

so build-time and reload-time agree.

Verified end to end against arXiv 1706.03762v7: build → LocalKnowledgeLibrary.put_paper → summary search now completes and returns a cited global summary.

Verification performed

• bash scripts/verify.sh — ruff format/check, basedpyright, lint-imports, pytest --cov: 491 passed, 86% coverage.
• New tests: whitespace-variant quote accepted and non-matching quote rejected, at both the knowledge layer (tests/knowledge/test_paper.py) and the flows research validation (tests/flows/test_paper.py).
• Manual end-to-end run of the build → persist → search path on arXiv 1706.03762v7.

🤖 Generated with Claude Code