v7.29.0

Engine identity: engine-service-v0.1.100 (was v0.1.98). Deployed and
cloud-gated before the floor moved; seven additive schema changesets.
v0.1.99 was cut and deployed first and then superseded the same day: its
intra-op thread policy capped the shared ONNX session at two threads and
made local indexing about 3.4x slower (nexus-00wsf); v0.1.100 leaves
ORT's default unless an operator overrides it.

Engine

• No SQL strings in Java (nexus-zrcj7): the vector search paths, the catalog
  FTS match, the shutdown reaper, the taxonomy sequence and centroid ANN
  query, and the migrator's bootstrap reads now run through generated jOOQ
  DSL or Liquibase SQL functions jOOQ generates tables for. Hybrid search is
  a probe (gate chashes) plus a by-chash cosine rank, with the HNSW-first
  dense-gate branch preserved as its own function; order and distance are
  byte-equivalent to the retired path. RawSqlGateTest now scans raw JDBC
  statements, DSL.sql, raw where-string overloads and SQL-bearing DSL
  templates, with a five-entry reduce-only exemption registry.
• Local embeds and reranks are admission-controlled process-wide (one
  shared gate, bounded interactive wait), and the ONNX intra-op thread count
  is an explicit operator override only, never derived (nexus-00wsf; the
  original "CPU spin" was attention-tensor thrash, fixed since v0.1.81).
• The engine release workflow creates the GitHub release as a draft and
  publishes it only after both build matrices succeed and all 21 assets are
  attached, so a tag can no longer exist without its binaries (nexus-cl14i).

Changed

• RDR-200 nx_answer continuation mode is parked: three ship-gate runs showed
  a session searching on its own beats the plan-plus-handoff. The opt-in
  code stays dormant; the headless path remains the default.
• Plugin skills using-nx-skills and orchestration reworded in plain
  register (nexus-ht9m5).

Fixed

• Fresh installs ran a MinerU no gate had seen. mineru>=3.1.11,<4 let a
  fresh resolution land on 3.4.5 (published 2026-08-14) while uv.lock and
  every test stayed on 3.1.11; the same shape holds for docling (2.76.0
  locked, 2.125.0 installed). The pin is now <3.2, the minor the fixtures
  are locked against; a lint ties that cap to the lock for shape-sensitive
  packages; the dependency drift watch reports any version change in that
  set, not only major jumps, and now actually runs weekly; and the developer
  reinstall constrains to uv.lock so the box matches the gates. Bumping
  MinerU or docling is a deliberate lock-and-cap change on a green slow-gate
  run. First such pass, same day: docling moves to 2.125.0 (cap <2.126),
  the version every fresh install has run since mid-August, gated by the
  unit suite, fresh-install MVV and shakedown; mineru stays at exactly
  3.1.11 (cap <3.1.12), because both candidates were gated and refused:
  3.4.5 collapses every "ff" ligature ("efficient" becomes "eficient", 67
  words to 0 in one fixture paper) and 3.1.15 wraps plain text as inline
  math and splits a word. The slow MinerU test now locks the ligature
  count. A PDF indexed on a fresh install between 2026-08-14 and this
  release ran 3.4.5 and carries the mangled words until re-indexed.
• PDF indexing dropped every table (nexus-jd8fi). pdf.mineru_table_enable
  defaulted to false, and the MinerU server was spawned with
  MINERU_TABLE_ENABLE exported, which MinerU lets override the per-request
  flag; every table came out as ![](images/<sha>.jpg) and no tabular value
  was indexed. Measured on the Pangram 4 Technical Report: pages 18-21 gave
  0 tables and 4515 chars off, 6 tables and 13890 chars on. The default is now
  true, the env var is no longer exported so each request's flag governs,
  and any remaining image reference (a figure, or a table MinerU could not
  read) is indexed as a caption-anchored marker such as
  [Table 6 not extracted as text; values not indexed] instead of an empty
  line. MinerU-extracted tables now populate table_regions, so their chunks
  carry chunk_type=table_page as docling's already did. PDFChunker is
  table-aware: a table that fits the window moves whole into the next
  chunk, an oversized one breaks after a complete row and each
  continuation chunk repeats the header row, so a cell value is never
  torn from its column name (real tables run 2200 to 5000 chars against
  the 1500-char window). Documents indexed before this need a re-index
  to pick up their tables.
• store_get_many cut long documents at the display cap with a bare
  ellipsis, so every tool-free reader downstream (nx_tidy, the plan
  runner's operators) reported healthy entries as truncated and asked for
  re-ingestion. The cut now carries a "TRUNCATED FOR DISPLAY, not a defect"
  marker at the point it happens (nexus-lugwx, generalising nexus-c0sdc).
• Plan runner: document tumblers returned by the metadata-, topic-, graph-
  and aspect-scoped searches hydrated to empty strings on both the operator
  auto-hydration path and an explicit store_get_many step; they now
  resolve through the catalog manifest. An operator whose evidence filters
  to nothing no longer dispatches a model call and returns a fixed
  "No evidence to reduce" result instead (nexus-mm5tx).
• Plan runner: parallel reduce steps with no fan-in were reduced, paid for,
  and discarded silently. They are now detected and reported in
  nx_answer's envelope as dropped_reduce_steps, with a one-line notice
  in text mode (nexus-4h0oh).
• nx_answer: the zero-evidence fallback now names the matched plan and each
  retrieval step's tool, resolved corpus and query, in the return and the
  telemetry row (nexus-ivv4d); the continuation text instruction carries
  collection and distance per evidence item (nexus-kim0o).
• nx answer-runs under-counted degenerate runs about 9x: the listing
  reroute, budget-warning-only and empty-hydration bodies are now
  classified, with error rows checked first and the reroute header anchored
  to the first line (nexus-x79ne).
• Operator dispatch: a child's cold start no longer consumes the processing
  budget (the timeout now runs in two phases, first output then the rest,
  bounded by an optional absolute deadline); both pipes are drained from
  spawn so a verbose child can no longer deadlock the dispatch, and post-kill
  cleanup is bounded rather than waiting on pipe EOF (nexus-tx5hd). A real
  killed-child test now proves the partial-output drain recovers streamed
  output (nexus-q4o43).