v3.9.3 — a slow database deserves a longer fuse

[3.9.3] — 2026-09-07 — a slow database deserves a longer fuse

Two fixes for the remote-database topology found while deploying 3.9.2 against
SurrealDB over wss.

Fixed

• SURREAL_MEMORY_MCP_TOOL_TIMEOUT (seconds) makes the MCP server's per-tool-call
  budget configurable instead of the hardcoded 30 s. A client talking to a remote
  SurrealDB can legitimately need more — recall runs embedding + vector search +
  rerank across the wire — and the fixed budget made smem_recall and
  smem_context time out for every desktop client while the same operations
  succeeded from the CLI. Values above 600 s clamp with a warning (a millisecond
  typo must not become a five-hour hang), invalid values warn and fall back to
  the 30 s default, and the effective timeout is logged at server start.
• ensure_schema converges a legacy SCHEMALESS change_log table before applying
  the field definitions: it probes INFO FOR DB and runs
  ALTER TABLE change_log SCHEMAFULL — the documented schemaless-to-schemafull
  transition, keeping existing rows and rejecting nothing the writer produces.
  Databases that wrote change-log rows before the table was declared in the
  schema (a shape an export/import preserves) previously warned on every startup
  and the payload field never converged to its declared FLEXIBLE form. Probe
  and ALTER are fail-soft: a failure preserves the previous behaviour exactly.