v3.4.0 — dedup honesty, native drift detection, and a 3D graph

Added

• Native SurrealDB drift detection. Tag-cooccurrence + Jaccard-similarity clustering —
  removed from the codebase along with the SQLite backend, leaving smem_uncertainty and the
  dashboard's Drift card silently reporting "0 clusters" indistinguishably from a real analysis
  that found nothing — is back as a first-class SurrealDB feature. New detect_drift
  consolidation strategy; the Drift card no longer claims the feature is "SQLite-only".
• 3D force graph. The Graph page moved from a flat 2D sigma.js layout to a 3D force-directed
  graph (particle flow along links, click-to-focus on a neuron's neighbourhood with camera
  fly-to). The node-count slider is bounded by what the API can actually return and its
  annotation names the real brain total, not a sampled count.

Fixed

• Dedup consolidation reported "0 new alias links" identically for "already fully linked"
  and "every attempt failed". The report now distinguishes the two, with machine-readable
  counters alongside the prose summary.
• Cloud Sync's Devices card always showed zero devices, and the local machine never
  self-registered — both were a silent AttributeError on fields DeviceInfo didn't carry,
  swallowed by a bare except.
• remember_batch under-reported success: a batch where every item failed validation was
  still counted as a success at the top level, because failure was only detected via an
  "error" key the handler never set. Also fixed: trust_score/source_id/context were
  silently dropped from the persisted item; a non-string content field crashed the whole
  batch instead of failing just that item.
• auto_capture only recognized English causal phrasing and truncated captured text
  mid-sentence rather than at the last sentence boundary.
• semantic_discovery always attacked the same ~26% of a large brain. Candidate selection
  read the first N neurons off a pagination order that never changes between runs; it now
  prioritizes the least-connected neurons first, so repeated consolidation passes actually
  reach the rest of the graph.
• A dashboard read of a non-default brain (GET /brains, GET /stats) could permanently
  repoint the shared storage connection to that brain, so a concurrent MCP call or background
  job racing with the dashboard could silently read or write the wrong brain's data. Reading a
  specific brain's diagnostics now goes through an isolated, scoped connection instead of
  mutating shared state.
• Deleting a typed memory could silently no-op when given the folded record-id form that
  find_fibers returns, rather than the dashed field form — the delete matched nothing and
  reported success anyway. Both id forms are now recognized.
• Consolidation's dormant-neuron replay scanned every neuron in the brain to build a small,
  bounded sample; the sample is now taken directly from storage instead of loading the full set
  into Python first.
• Keyword document-frequency counts could silently stop updating. A formatting bug in a
  batched write statement made every count update a no-op instead of an increment; a second,
  related issue could make the write collide with a legacy record id left over from an old
  brain rename. Both are fixed — the write now matches by keyword content instead of by a
  computed id, so it can no longer collide with a stale id shape.
• remember's content-length limit could be bypassed by supplying enough extra context that
  the combined, merged text exceeded the configured maximum — the length check ran before the
  context merge, not after.
• Dev tooling: the project's Python interpreter version was unpinned, so mypy parsed
  third-party type stubs (e.g. a newer numpy's) against whatever Python version uv happened to
  resolve on a given machine, rather than the version this project targets. A .python-version
  file now pins it.

Changed

• CI/release: the dashboard's unit-test suite now runs before the dashboard build, in both the
  CI and release pipelines — previously nothing ran it.