v3.3.0

Added

• A CI job now runs the full test suite against a live SurrealDB. Every prior job
  either mocked the database or skipped whenever SURREALDB_URL was unset, so the only
  production backend went untested in the pipeline — the exact gap #143's bug walked
  through (a query that returns plausible garbage instead of failing loudly). The job
  starts the official image via a docker run step (GitHub Actions' services: cannot
  pass the datastore/flags SurrealDB's entrypoint requires), waits for /surreal isready,
  then runs the whole suite sequentially — no -n auto, since parallel workers sharing one
  connection produce transaction-conflict errors unrelated to any real regression.

Fixed

• Polish stop words ma/na/co/sa swallowed the acronyms MA/NA/CO/SA.
  A contributor's audit (#64) showed the fix everyone assumed — removing the words from
  the stop list — was wrong: N/A and S.A. never actually collide (punctuation
  fragments them below the minimum word length before the stop-word check ever runs), so
  removing the words would only trade real function-word noise for a niche acronym gain.
  A token that is ALL-CAPS in the source now survives even when its lowercased form is a
  stop word, gated by how much of the text is uppercase — a shouted note or an all-caps
  heading has no acronym to rescue, only every word incidentally capitalized, and must not
  have its stop-word filtering disabled wholesale.
• check_dead_modules passed silently while five files imported a module that no
  longer exists. Reachability is computed from each file's imports, so one naming a
  module the tree no longer has (the SQLite backend, removed in 3.0.0) contributed
  nothing to the graph and produced no diagnostic — the guard reported No unreachable modules. while five benchmark/script files could not run. It now resolves every import
  against the tree and reports the ones that don't. The five files were fixed: one
  (stress_at_scale.py) measured only the removed backend and is deleted; the other four
  now require a live SURREALDB_URL instead of silently falling back to it.
• Two read-only endpoints in the sync hub could redirect scheduled maintenance onto
  the wrong brain. GET /hub/status/{id} and GET /hub/devices/{id} switched the
  process-wide shared storage's active brain to answer a lookup, and the background
  consolidation/decay loops read that same mutable state on their next tick — so a
  read-only request for brain B could cause the next scheduled pass to run against B
  instead of the brain the operator actually left active. Both endpoints now read through
  an isolated, brain-scoped connection instead of mutating the shared one — the same
  pattern already used for reasoning-training's read endpoints.
• The tool-stats days filter changed the daily chart but not the summary above it.
  get_tool_stats took no days argument, so its summary was always computed over all
  time; switching the dashboard's range selector between 7/30/90 days produced a
  byte-identical summary while the chart below it changed correctly. The summary now
  respects the same window as the daily series, and the two storage methods behind it are
  declared on the storage interface (with an in-memory implementation) instead of being
  reachable only on the SurrealDB backend behind an unchecked attribute access.
• A SELECT VALUE query on an array-typed field could be read as if it were rows.
  The shared query helper's return type says it hands back row dictionaries, which is not
  what a SELECT VALUE query returns — the field it selects for one matching row can
  itself be an array, indistinguishable by shape alone from several separate scalar rows.
  That mismatch is the mechanism behind a prior release's fixed bug (a result iterated
  character by character). The one live call site reachable this way now goes through a
  separate, honestly-typed helper instead of the row-shaped one.
• The Settings brain-files panel showed a plausible path to a file that was never
  written. A brain that exists only in SurrealDB has no on-disk database file, but the
  panel built and returned a path for it regardless — worse when a stale file from an
  older install happened to sit at that exact path for one brain while its neighbours
  showed nothing. The path is now omitted rather than fabricated when the file does not
  exist.

Removed

• The smem_drift tool and its underlying tag-drift detector. Every call path into it
  swallowed its own failures, so on the only shipped backend it always reported "clean" —
  indistinguishable from a real analysis that found nothing, because the two things it
  measures (tag co-occurrence, session summaries) were never implemented on that backend
  and never had been, even before the SQLite backend's removal. Its merge action never
  actually merged tags either. TagNormalizer's own drift detection, which needs no
  storage and already backs smem doctor's tag-drift warnings, remains and covers the
  same practical need. A small unrelated write path — recording tag co-occurrence for the
  now-removed detector, and a periodic session-summary persist with the same swallowed
  failure — is removed alongside it.
• A CLI sandbox-guard helper was renamed from ensure_aiosqlite_or_exit_cli to
  ensure_sqlite_or_exit_cli: it has only ever checked the stdlib sqlite3 module, never
  the optional aiosqlite package, and the name was misleading anyone reading the CLI
  startup path. Internal-only; no public interface changed.

Changed

• BREAKING: the smem_drift MCP tool is gone. If anything called it expecting a real
  answer, that answer was never trustworthy on the SurrealDB backend to begin with — see
  Removed, above.

Full diff: v3.2.0...v3.3.0