fix(memory): separate uploaded documents from facts

[aliyevaladddin]

A spreadsheet upload was landing in shared memory as 29 "facts". The dashboard showed Awards ( 6 hours) NaN NaN 2.0 ... Total before Taxes NaN NaN NaN 34980 under "Recent Shared Memory", and the same slices were injected into every agent's system prompt via build_shared_context_block.

Three defects, one cause — a document chunk was being treated as a fact:

• df.to_string() is a console rendering. It emits "NaN" for blank cells, which an LLM reads as a value, and truncates wide frames. Now to_csv(na_rep=""), across every sheet rather than only the first.
• Chunking sliced on character count, so rows were cut mid-number and the header appeared only in chunk 1 — every later chunk was unlabelled figures. Tabular text now splits on line boundaries with the header repeated on each chunk.
• Chunks were stored with visibility="shared", so they were injected always and crowded real facts off the dashboard. They now go to a document_chunks collection, retrieved on demand through a new search_documents tool.

A fact asserts something ("Acme renewed in March"); a document chunk is raw source text with no subject. They are different kinds and now live apart.

Existing uploads are not deleted: fact reads exclude the file-upload tag, and POST /api/mongodb/documents/migrate-legacy moves them across, reusing their embeddings so it costs no provider calls. Filtering happens after the vector search rather than in the $vectorSearch filter, because Atlas only filters on indexed fields and tags is not one — doing it there would error into the recency fallback and lose ranking.

Also adds GET /documents, GET /documents/search, DELETE /documents/{filename}. The new collection needs a vector index created by hand in Atlas; until it exists, search falls back to recency so uploads stay reachable.

Frontend: @floating-ui/utils pinned and lockfile updated (an interrupted npm install had left it with an empty dist, so @base-ui/react could not resolve it).

Verified: 250 tests pass (7 new). Removing the legacy filter fails exactly the two injection tests, so they pin this regression rather than passing vacuously. Extraction checked against a sheet with blanks and a totals row — no NaN in the output; 120-row table chunks into 8 parts, every chunk headed, every row intact and in order.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.