v1.12.0

Memory gains an inbound path: vectr memory import reads existing on-disk
memory files into the store, so adopting vectr no longer starts with retyping
what you had already written down. vectr_unanchor completes the anchor pair
and brings the MCP surface to 23 tools. The rest of the release is retrieval
work, most of it the result of measuring things that had been asserted:
a similarity metric that was reading almost the inverse of the truth on
repetitive code, a dedup key that could not fire on the case it was built for,
and a guard whose rule only worked on same-length renames.

Working memory

• vectr memory import brings existing memory files into the store. vectr
  could write memory out and never read it back: vectr memory export renders
  notes to markdown and nothing parsed that file back, and there was no path in
  from anywhere else, so a user arriving with a MEMORY.md, CLAUDE.md,
  AGENTS.md, .cursorrules or copilot-instructions had to retype all of it
  one call at a time before the tool had earned any trust. Import discovers
  those files, parses at Markdown heading granularity, and writes notes.
  --dry-run prints the plan and changes nothing. Re-importing an unchanged
  file creates zero notes; an edited section creates a new note without
  deleting the old.
• Provenance on import is the load-bearing decision, not a detail.
  remember() rejects provenance='auto' on kind='directive', since an
  unreviewed standing rule is a contradiction in terms, and an imported rule is
  exactly that pair. Foreign imports therefore land as auto, except imperative
  rules classified as directives, which land as agent: the lightest class that
  clears the guard without claiming an endorsement nobody gave. The user
  promotes after review. Re-importing a file vectr itself exported is a separate
  path that preserves the provenance the export recorded.
• vectr_unanchor removes anchor paths from an existing note, the inverse of
  attach, across REST, MCP dispatch and schemas, and the CLI. It is a plain
  update plus one audit call rather than a lifecycle event, because anchors are
  delivery-shaping metadata and staleness is re-derived from live hashes on
  every check. Idempotent, exact-string match on the path to mirror attach,
  ValueError on empty anchors, None for an unknown note, and an empty list
  rather than null when the last anchor goes.
• Kind-filtered recall no longer under-fills silently. _semantic_recall
  applied the kind filter after a fixed limit * 3 fetch, so a selective kind
  could return fewer than limit with no signal that the shortfall was a pool
  artifact rather than an absence of matching notes. The pool widens when kind
  is set, and a metadata-only prefetch backstops the relevance branch, its
  survivors appended after the ranked pool.
• sort_by is one shared vocabulary rather than four. Dead states no longer
  count toward the stale-task total.

Retrieval and ranking

• Near-duplicate dedup compares bodies with the leading docstring stripped. The
  near_dup_body key existed for the case where two chunks carry different
  docstrings over near-verbatim bodies, and it could not reach that case at all:
  it compared whole chunks, so the differing docstrings were inside the
  comparison and dragged the ratio below the threshold. On the feature's own
  witness this moves similarity from 0.765 to 0.958.

• A similarity defect was found underneath that, and it is the more important
  finding. Every ratio in this area was computed with difflib's default
  autojunk, which excludes any element appearing in more than one percent of a
  200-element-or-longer sequence. These are character sequences. Measured:

  repetitive body, 392 chars    0.166 default    0.995 corrected
  varied body,     251 chars    0.996 default    0.996 corrected
  repetitive body, 150 chars    0.993 default    0.993 corrected
  

  So it needs both length and heavy repetition, and on ordinary varied code it
  changes nothing. The content it distorts is repetitive text, which is
  templated code, which is also the false-collapse hazard for these keys, so the
  broken metric had been acting as an accidental guard.

• autojunk=False is therefore applied only where it ships no behaviour: the
  disabled near_dup_body key and its guard. The enabled docstring key keeps
  the wrong parameter deliberately, recorded at the call site, because its 0.75
  threshold was calibrated through the same distortion and correcting the metric
  alone collapses 50 templated accessors into 1 result.

• ranking.near_dup_body ships disabled pending measurement. The threshold's
  safety margin was argued from a single witness pair and does not survive
  templated code, where bodies differing only in an index cross it while being
  genuinely distinct symbols.

• The templated-body guard is now alignment-independent. The previous rule
  walked SequenceMatcher opcodes and rejected any pure insert or delete, so it
  only ever worked for same-length renames: name against phone is four
  characters against five, forcing an insert, and the rule rejected the very
  shape it existed to catch. It now masks identifier-like tokens and bare
  numbers to one placeholder and compares the masked forms.

• Wiring that guard into the enabled docstring key was tried and reverted,
  which settles an open question. That key exists because three scripts sharing
  a copied docstring crowded out a canonical struct definition, and those three
  differ only in an identifier, exactly like the accessor family that must never
  collapse. The must-collapse witness and the must-not-collapse hazard are the
  same shape, so no rule reading the characters of two bodies separates them.
  Correcting autojunk there needs a signal this layer does not have.

• A relative-cliff sub-signal for the low-confidence banner ships behind a
  switch that defaults off, with a distribution harness to calibrate it. The
  existing absolute floor was tuned against the previous cross-encoder, so the
  swap left the constant stale, and neither raising it nor replacing it with a
  rank-1-to-rank-2 test is safe on the evidence that exists.

Symbol graph

• Method-override edges are recorded at index time. A method whose class's base
  declares the same name now emits an overrides edge, reusing the existing
  edge type so no schema migration was needed. Only simple-identifier bases
  resolve; an attribute-lookup base is skipped rather than guessed, because a
  wrong override edge would later push the wrong symbol up. overrides_of and
  overridden_by read it back.
• JavaScript and TypeScript produced zero override edges because the extractor
  looked up class_heritage as a field when it is a child node type, so the
  lookup always returned nothing. Traversal now scans children, reads the
  identifier directly under it for JavaScript, descends through
  extends_clause for TypeScript, and skips implements_clause, since an
  interface method has no body to override.
• Java enum constants and fields are indexed. The impl error check is scoped.

CLI and editor integration

• Command normalisation keeps escaped separators unsplit, pads glued single
  pipes, and handles a backgrounding single ampersand.
• The README's MCP tool count was wrong in three places including the section
  heading whose anchor the badge links to, so the badge pointed at a heading
  that no longer existed.

Benchmarks and internals

• The acceptance corpus can run over the MCP surface. It drove REST only, so it
  was structurally blind to the two things that decide whether a caller receives
  an answer: the low-confidence banner and pointer mode. A case could be scored
  passing with the expected symbol at rank 1 while the caller was told to go
  grep. low_confidence_absent and body_present assertions are meaningful
  only on MCP and emit an explicit skip notice under REST rather than passing
  silently. --strict-status compares each case's recorded label against the
  observed outcome and fails on drift, which is how four labels had drifted
  unnoticed.
• The six-task sprint harness can drive a routed model, so its numbers can be
  re-measured on something other than the model they were first taken from.
  Results carry a scope stamp naming the model, route, date and revision, an
  unparseable metric is recorded as absent rather than zero, and upstream
  failures are recorded per task rather than retried away.
• New harnesses for the banner threshold distribution and the docstring-key
  similarity threshold, plus an inventory of every measurable README claim with
  its source and method.
• Test isolation fixes: the instance registry is isolated, the dummy embedder is
  deterministic rather than salted by hash seed, and a benchmark module rename
  ends a sys.modules collision where two test files imported a bare harness
  from different directories and whichever was collected first won.