fix(ai-service): grounding verifier rejects quotes that span a table cell border #50

Beobachtet

Eval case t03-table-eml-plaintext (#24 / PR #41): the model quotes a table row across a cell border, e.g. 60 | Stk.. The grounding verifier does not confirm the unit from that quote, so a correct value ends up not found.

Erwartet

A quote whose text matches the cited segment once cell separators (|, tabs, repeated spaces) are normalised is confirmed – as long as the value itself is in the quote and in the cited segment. "Found needs proof" (ADR-0001 D8) must stay strict: no fuzzy matching beyond separator normalisation.

Reproduktion

1. uv run --project services/ai --directory services/ai python -m requestflow_ai.evals --replay
2. See line_items.unit / line_items.quantity for case t03: the value is not grounded.

Akzeptanzkriterien

• Fehler behoben, Regressionstest existiert (unit test in the verifier tests with a cross-cell quote).
• Negative test: a quote that only matches after dropping real characters (not separators) is still rejected.
• Eval replay: t03 grounded; no key field drops; baseline updated deliberately.

Security/Privacy betroffen?

KI/RAG: the verifier is the proof gate for found – change needs the eval gate and a review against D8.