v0.4.18

Patch release. The theme is evidence: a drawing can now be asked what it recognised,
what it missed, and what it did with each finding — and say so in versioned JSON that
another program can read.

Recognition evidence and reporting

• inspect_step(path) (#1460) — read-only, versioned STEP evidence over one hashed
  byte snapshot. No drawing build, view projection, annotation placement, render, export
  or lint path runs; a guard pins the exact set of engine modules an inspection executes.
  It reports what was found, what each finding was used for (acted_on, disposition,
  reason, owners), and what went unclaimed.
• Report v1 and CLI sidecars (#1438, #1449, #1455, #1456) — a versioned
  machine-readable recognition and drawing report written alongside generated artifacts,
  atomically (#1452), with physical requirement outcomes projected into it.
• Script generation writes the sidecar to disk (#1454) rather than embedding a
  snapshot in the emitted Python.
• Occurrence ownership (#1441–#1453) — every accepted recognition occurrence is now
  tracked to an outcome: direct, grouped, nested countersinks, conditional channels,
  turned steps, through steps, round bosses, exact Plate ownership, and an explicit
  classification for ownerless occurrences.
• b123d-recognisers 0.4.14 Blend paths adopted (#1459).

Fixes

• A raw OCC failure no longer escapes on Linux (#1395, #1468). A degenerate iso zone
  reached gp_Trsf::SetScale and crashed: on Linux the kernel raises
  Standard_ConstructionError, which is not a Standard_Failure subclass, so the
  handler missed it. Two further cases came out of the same fix — a negative factor was
  accepted silently and mirrored the geometry, and a NaN scale slipped past a <= 0
  test and hung the build.
• *_requirement_unverifiable no longer prints ? where a requirement id belongs
  (#1397). That finding is recorded exactly when a recognised source matched no IR
  feature, so no id exists; the message now names how many requirements are unattributable,
  or says the cardinality is unknown where it genuinely is.

Documentation

• The ADR corpus is five records (#1463), down from twenty-one and 8,787 lines, each
  capped at 200 lines and each naming the guard test that fails if an invariant is
  violated. The superseded records are archived as frozen history.
• The multi-feature object-reference workflow doc now runs (#856, #858). Its example
  raised ValueError as written, and a test now executes what the doc tells the reader to
  execute and pins the measured output.

Internal

• reporting's occurrence projector is a named seam (#1461): project_occurrences,
  json_value, producer, write_json_document, drawing_report.
• PR CI shards were 1.60x out of balance — 38% of collected tests had no recorded
  duration. Rebalanced to 1.00x, cutting PR wall clock by ~19%.

Full changelog: v0.4.17...v0.4.18