fix(lancedb): tag search span as RETRIEVER + populate input/output

[SuhaniNagpal7]

Summary

LanceDBInstrumentor emits a span for table.search() (.to_list() / .to_pyarrow() etc.) but never sets the FI canonical retriever keys. Future AGI dashboard renders the span as Type: unknown with empty Input/Output panels.

LanceDB uses a query-builder pattern where the actual query lives on the builder instance as _query / _limit (instead of being in args/kwargs), so the wrapper reads from those attributes.

What changes

All in traceai_lancedb/_wrappers.py:

• Optional fi_instrumentation.fi_types import with raw-string fallback.
• SearchWrapper.__call__:
  • gen_ai.span.kind = \"RETRIEVER\"
  • Input summary {limit, output_format} extended with either query (capped at 500 chars, for text queries) or vector_dim (for vector queries) based on what's on the builder instance.
  • input.value = JSON of the summary; input.mime_type = application/json.
  • After the wrapped call, output.value based on output_format:
    • to_list → JSON of the first 50 rows
    • PyArrow Table (has num_rows) → JSON via to_pylist() or to_pydict() of the first 50 rows
  • output.mime_type = application/json in both cases.

Add/Update/Delete/CreateTable/DropTable/OpenTable are untouched.

Verified

• In-process attribute check using a mock query-builder instance with _query=[0.1,0.2,0.3] and _limit=2.
• Real end-to-end ingest to Future AGI → confirmed via Future AGI MCP that the lancedb search span shows:
  • Type: Retriever
  • Input: {\"limit\": 2, \"output_format\": \"to_list\", \"vector_dim\": 3}
  • Output: [{\"id\": \"doc1\", \"text\": \"Eiffel Tower\"}, {\"id\": \"doc2\", \"text\": \"Statue of Liberty\"}]

Out of scope

Per-document retrieval.documents.N.* attrs (Tier 3). Other 6 vector DBs get their own PRs.