Heal LLM-simplified citations to canonical reference

[John CSA (jluocsa)]

Purpose

Fixes #2950.

When a document filename contains leading special characters (e.g. a leading
hyphen, as produced for some user-uploaded files: - PyCon US 2025.pdf), the LLM
tends to "simplify" the citation it emits, dropping the leading characters. For
example, given the source - PyCon US 2025.pdf#page=1, the model writes
[PyCon US 2025.pdf#page=1] in its answer.

The frontend citation parser validated the cited text with endsWith, which
correctly recognized it as a valid citation, but then used the simplified text
as the citation reference. The resulting /content/<path> request used
PyCon US 2025.pdf instead of the indexed - PyCon US 2025.pdf, so
check_path_auth() filtered by the wrong sourcefile value and the document
returned 403/404.

This change resolves the cited text back to the canonical citation from
data_points.citations (exact match preferred, otherwise the endsWith match)
and uses that canonical value for the reference and for the
citation_activity_details lookup. Links and lookups now use the indexed value,
so the content route resolves correctly.

Does this introduce a breaking change?

[ ] Yes
[x] No

Citations that already match exactly continue to resolve exactly (the exact match
is preferred). Only previously-broken simplified citations change behavior.

Does this require changes to learn.microsoft.com docs?

[ ] Yes
[x] No

Type of change

[x] Bugfix
[ ] Feature
[ ] Code style update (formatting, local variables)
[ ] Refactoring (no functional changes, no api changes)
[ ] Documentation content changes
[ ] Other... Please describe:

Code quality checklist

• tsc type-checks cleanly (npx tsc --noEmit).
• prettier applied to the changed file.
• The current tests all pass (python -m pytest).

[John CSA (jluocsa)]

Updated this PR to implement the maintainer's suggested Solution #1 (post-process the LLM response) server-side, in addition to the original frontend safety net.

What changed

• app/backend/approaches/approach.py — added heal_citation / heal_citations helpers that map a simplified citation token (e.g. [PyCon US 2025.pdf#page=1]) back to the canonical indexed reference via exact match, else a unique endswith match. Ambiguous or no match leaves the token unchanged.
• app/backend/approaches/chatreadretrieveread.py — applies healing in both the non-streaming and streaming paths (including citations split across token deltas).
• app/frontend/src/components/Answer/AnswerParser.tsx — original frontend fix retained as a display-layer safety net.
• tests/test_chatapproach.py — 10 new tests covering exact/unique/ambiguous/no-match cases and a citation split across streaming deltas.

This means corrected citations are now present in the raw response itself (streaming, API consumers, exports), not just the rendered UI. Rebased onto current main; all healing tests pass and ty is clean.