The reconcile pass left one real defect for the reviewer. Groups U and N are
built side by side, so the takeoffs carry their own unit bands, cluster grid,
scale statistic and room faces, and nothing ever retires them. Two readers of
one drawing can disagree. On Task 6's own inches-over-millimetres layout,
infer_units without room faces ties three ways and falls back to the header,
while the takeoff's bands say mm.
Task 11 gains Steps 15-20, committed on their own before the README figures
are retaken:
- describe.drawing_units becomes the one unit call, used by describe,
scale_check and the takeoffs;
- the takeoffs read plan copies, scale and rooms through group U's modules;
- the stand-ins are deleted;
- tests/test_understand_one_reading.py pins one unit, one plan copy, one
verdict and one set of rooms per drawing.
Tasks 3, 4 and 6 now point at those steps. Review Focus 1 names
parse_diameters, the function every task uses.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FKFStjEFzpGyLq5bMpWxTd
A second pass compared every tag, number and literal in the plan with the
field files, locally. Two borrowed values remained: a motor rating and one
motor tag. Both are now invented (18,5 kW, M82), and the dependent sums and
section rules were checked against them. The rest of the tank series was
renumbered to keep its old sort order, since several assertions compare
sorted tags: T4101-T4103, T4500 and T4600.
Docstrings no longer tie example layer names to a particular drawing. The
fixture's claim now says exactly what is invented (every tag and number) and
what is ordinary drafting vocabulary (layer and block names). The spec's
grounding paragraph now says "process plant" and no longer names the sector.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FKFStjEFzpGyLq5bMpWxTd
The skeleton's header, constraints, file map and pinned interfaces now carry
their twelve tasks, in order:
- main (Tasks 1-2): the snapshot, vocabulary and label parsers, MAX_DXF_BYTES
512 MB and the office extra; the synthetic plant pair with its truth.
- U (Tasks 3-4): drawing_understand and drawing_scale_check; SECTION 26 opens.
- N (Tasks 5-8): the line network, orthogonal lengths and pipe rows, the
takeoff workbook with pipe_takeoff, then cable_takeoff; SECTION 27, pack plant.
- D (Tasks 9-10): drawing_diff and drawing_topology_check with the three
topo_* focuses, on D's own copy of SECTION 26.
- final (Tasks 11-12): packs, lean 65, aliases, golden queries and counts;
benchmark v7, five correctness checks, the live smoke, the field-test
script and the docs.
Interface pass: the U and N texts now say what Task 6 really does with the
scale check (it takes one through pipe_rows(scale=...) and otherwise keeps its
own stand-ins), and Task 11 moves D's block above SECTION 27 after the merge.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FKFStjEFzpGyLq5bMpWxTd
Header, global constraints (confidentiality, read-only by default, nothing
invented, inferred units), the five real-world inputs no happy-path test
exercises (AutoCAD text codes, Cyrillic look-alike tags, repeated tags,
double-drawn pipe, a layout whose INSUNITS lies), file structure, the
interface block, execution with the worktree fast-forward rule, and the merge
notes. Task sections are drafted next.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FKFStjEFzpGyLq5bMpWxTd
