fix(lancedb): translate the spill failure on the write path too

[gloryfromca]

Summary

#463 translated LanceError(IO): Execution error: Spill has sent an error into the retryable VectorStoreBusyError only inside the read deadline wrapper. The soak that ran with #463 shows that was the wrong place: 42 spill failures, 0 translated, 104 rows filed unrecoverable. The traceback frames (Windows soak, everos_soak_run3_integration883007c/server.log):

worker.py:800  outcome = await handler.handle_added_or_modified(...)
  -> _apply_lance_changes -> index_repo.upsert(to_upsert)
  -> router.py:48 -> backends/lancedb.py:115 -> repository.py:466 (merge_insert)
  -> lancedb/table.py:5269 execute_merge_insert
RuntimeError: lance error: LanceError(IO): Execution error: Spill has sent an error

So it is merge_insert under _locked. Same translation there; the write is idempotent by id, so the worker's retry (backoff, ExternalServiceError path) is the right answer. Marker docstring corrected (four runs: 187 / 180 / 34 / 42).

Area

• architecture (storage)

Verification

• New test: the phrase raised inside _locked becomes VectorStoreBusyError; against main's _locked it fails (bare RuntimeError propagates).
• test_core/test_persistence/test_lancedb + test_cascade green.
• Live check: the next 2 h soak on main; lancedb_transient_execution_error must be > 0 and cascade_worker_unrecoverable back to the fuzz baseline (ValidationError count).

🤖 Generated with Claude Code