design.tex carries the contract — invariants, algorithms, public API,
roadmap, the canonical D1..D19 decision index, and a new System
Walkthrough (§17) that derives the geometry from first principles for
a graduate-engineering reader: vector rotations and the inward-normal
formula, the shoelace signed area, segment intersection via Cramer,
Sutherland–Hodgman half-plane clipping, the DCEL next-pointer rule
with worked square example, per-edge depth-cap ray-cast, corner-parcel
construction in two flavors, frontage walk, the shared-vertex
registry, propose-then-apply deform, and end-to-end traces for both
subdivide_all and apply_road_edit(MoveNode).
journal.tex carries the live record — Self-Decisions Checklist at the
top, sessions 1-4 rewritten condensed with explanatory Rust snippets,
and a Spec Deviations Log indexed chronologically.
Makefile builds both PDFs (`make all`, `make design`, `make journal`,
`make watch-design`, `make watch-journal`, `make figs`).
Old single-file journal content lives in git history per user.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The first half of the eventual full DCEL — vertex identity. Every
parcel polygon vertex now resolves to a stable VertexId via a
spatial-hash lookup at EPS_GEOM resolution. Coincident positions
across two parcels resolve to the same VertexId, so adjacent
parcels share one physical vertex.
ParcelSet::move_vertex(vid, new_pos) updates the registry's
position AND writes through to every parcel's polygon at the
recorded index. Adjacent parcels' shared boundaries can never drift
apart — they are the same vertex, mutated once.
The deform pipeline now propose-then-apply: deform_parcel_after_road_move
returns proposed (VertexId, new_pos) moves rather than mutating in
place. The outer loop validates each parcel's hypothetical
post-move polygon, then applies all proposed moves via
move_vertex. Conflicting proposals on the same vertex are
last-one-wins, but in practice the deform parameterization makes
all referrers agree by construction.
New regression test: shared_vertex_no_drift_under_repeated_edits.
50 small random node moves plus an inverse; asserts every shared
boundary vertex is bit-for-bit identical across every parcel that
references it.
Edge identity (parcel-layer half-edges) is the next milestone — it
enables split/merge ergonomics. Vertex identity alone is enough
for the no-drift contract this session was scoped to deliver.
Journal §11 session 4 entry adds D17 (registry), D18
(write-through), D19 (propose-then-apply), and the milestone-0.5
queue.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wrote real test bodies for the last two ignored tests using polyline
approximations of their curved geometries. Both verify I1-I3 hold
and don't panic. True pie-slice subdivision and proper tight-radius
depth clamping are still milestone-0.4 work, but the contract from
spec table tab:degenerate is satisfied — every named test exists,
runs, and passes.
Final count: 24 unit + 22 integration + 1 doc test = 47 tests, zero
ignored. Up from 14 active out of 21 named at the end of milestone
0.1.
Journal §11 session 3 numbers updated.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Y-intersection had a real I3 overlap at acute corners, caught by a
new programmatic centroid-in-other-polygon test
(y_intersection_no_overlaps). Fix: bisector-clip regular parcels
adjacent to acute corners (interior < 60°). Obtuse corners keep
their rectangle/parallelogram corner parcels and need no clip.
Minimum-change deformation: when a road's *line* doesn't change —
only its endpoints shift along the same line, e.g., the bottom road
gets longer when its right endpoint moves outward — parcels whose
frontage is still on the new segment are reported as Untouched and
keep their absolute coordinates. Only parcels on a road that
actually rotated get re-projected. Trade-off: vertex-exact
inverse-restore is no longer guaranteed (centroid drift bounded by
edit delta is the new contract).
SplitSegment preserve: rebind frontage_road for parcels whose
frontage is entirely on one side of the split point, or split into
two parcels along a perpendicular through the split point for
parcels that span. road_split_preserves test is now active and
passing. acute_intersection_15deg/5deg also active.
Test status: 24 unit + 20 integration + 1 doc passing; only
cul_de_sac and curved_road_high_curv still ignored (need real curved
roads). Journal §11 session 3 entry added with D14, D15, D16, two
spec deviations, and the milestone-0.4 queue.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Tracked files dropped from 11,756 to ~40 by adding .gitignore and
removing target/, *.aux/log/etc, and .DS_Store from the index.
Code changes since the initial commit:
- Milestone 0.2 (corner parcel rework, sticky back edges,
preserve-on-deform pipeline, performance instrumentation, new
Y-intersection and edit before/after figures, journal §11
session 2 entry).
- Milestone 0.3 in progress: I3 overlap fix at acute corners
(programmatic test caught a real I3 violation in the Y figure;
fixed by bisector-clipping regular parcels adjacent to acute
corners), minimum-change deformation (parcels on a road whose
*line* didn't change are skipped — only parcels on a road whose
direction changed are deformed).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
