v7.57.0

Pairs with engine-service-v0.1.129, unchanged from 7.56.0 — this release carries no engine-side change.

Fixed

• store_get no longer returns duplicated text when rebuilding a
  PDF-derived document. The whole-document rebuild joined a document's
  manifest chunks with no de-overlapping, which is correct for a store_put
  note (whose pieces do not overlap) and wrong for every PDF, because the
  chunker overlaps its chunks by 20% of the window by design. On an 11-page
  paper that duplicated 18.7% of the text and spliced words at the seams.
  Stored data was never affected — only the rebuild — so this repairs existing
  documents with no re-index and no user action. Affects the MCP store_get
  tool and nx store get alike. Measured across the live store: 134 of 140
  PDF-chunked documents, 30.4% of all knowledge documents.

• PDF table extraction flags a table whose rows disagree with its header.
  MinerU's table HTML was stored verbatim and checked by nothing, so a table
  whose header collapsed into a single colspan cell could read back stating
  the opposite of the source, with every value row left unattributable to a
  column. Two signals now mark such a table in place: a header whose width
  disagrees with its data rows, and a row that lost its label to a merge. The
  table's own HTML is never edited, and the marker rides inside the table so it
  reaches every chunk of one that spans several. Markers are added at index
  time, so they appear on newly indexed documents only.

• Path-keyed catalog writers refuse, report or announce instead of guessing.
  Several catalog documents per file_path is a normal steady state — nothing
  forbids it, the only uniqueness is the partial index on
  (tenant_id, source_uri), and it arises whenever one file is catalogued
  under two owners. The resolver already said when it chose among several; the
  callers did not. nx dt index now returns without stamping when the path
  names more than one document, naming every candidate, rather than stamping
  the first — writing identity onto the wrong row is how one run stamped the
  wrong document. The since-head delete reports the foreign-owner row it
  deliberately leaves live instead of skipping in silence. nx catalog session-summary shows every document for a path, where it previously
  reported one row's RDR links as the path's and could print "No linked RDRs
  found" while the links sat one row away. And the four single-document
  registrars announce before minting a second document for an
  already-catalogued path, recorded into a per-run collector that nx index
  resets and surfaces at the end of a run.

• The documented install command names its interpreter (nexus-sa187).
  uv tool install conexus is now uv tool install conexus --python 3.12
  everywhere a user is told to type it: README, the site's two install flows,
  docs/getting-started.md, docs/cli-reference.md, docs/configuration.md
  (both its "recommended setup" line and its [local] variant), and the
  nx-missing hint both preflight copies print. Ubuntu 26.04 LTS ships CPython 3.14 as
  python3, and on a box with no other interpreter uv picked it and the
  resolve failed outright — the torch pin (>=2.8,<2.9) has no cp314
  wheels. Every fresh install on the current Linux LTS hit it; measured
  2026-09-21 on WSL2 Ubuntu 26.04.1, and reproducible offline with
  uv pip compile --universal --python-version 3.14. Widening the torch pin
  would be a move toward 3.14, which this release cannot run, so the fix is to
  stop leaving the interpreter to the ambient python3.

  tests/e2e/fresh-install-mvv.sh --published had been passing --python 3.12
  all along while the docs passed nothing, which is exactly why the right gate
  at the right layer never saw this: the interpreter was the uncontrolled
  variable. tests/test_documented_install_interpreter_pin.py now holds the
  README, the site, docs/configuration.md and that MVV leg to one
  interpreter, and checks it against requires-python.

Added

• find_all_by_file_path on the CatalogReader protocol. The honest shape
  of a path lookup: a path can name several documents, and find_by_file_path
  returns only the first. Promoted onto the caller-facing contract because
  three production call sites now depend on it.

Removed

• conexus/hooks/scripts/preflight.py, the pre-RDR-215 implementation of
  the SessionStart preflight. hooks.json has named the nx-hook preflight
  verb since nexus-q02nx.21 and the /conexus:nx-preflight command goes
  through nx command-context, so nothing ran this copy; deleting it is that
  port's remaining half, recorded as the exit condition in
  tests/hooks/test_preflight_verb.py's own docstring and left undone when the
  port landed. It surfaced because the install hint it carried had to be
  hand-edited in parallel with its twin for nexus-sa187 — a duplicated fact
  with no consumer. The verb's tests drop the dual-implementation fixture they
  only needed while both existed; tests/test_nx_preflight_hook.py keeps the
  half that is about hooks.json rather than about an implementation, namely
  that the preflight runs above the guidance it counter-signals.

Known issues

• A document indexed from markdown (docs__, rdr__) can still read back with
  duplicated text at chunk seams. Its chunks record abutting spans, so the
  de-overlapping above does not apply to them. The cause is not established and
  a re-index may or may not resolve it, which is why this release promises
  neither. Tracked separately.

• A re-index after this release is not a no-op for a PDF with a flagged table.
  The marker changes extracted text, so that table's chunks get new content
  hashes. Expected for a text change, but worth knowing before re-indexing a
  large corpus.