feat(core): junction regeneration learns turn-set changes (p2-s2)

[JArmandoAnaya]

Closes #263. Highest-risk sprint in P2; gates #218 (Lane Carve).

What this does

regenerate_junction refused any turn-set change (twice — connection count, then per-turn key) and the editor swallowed the failure into a warning toast, so adding a lane to a road feeding a junction left it stale until a manual delete-and-recreate. That is exactly the case GW-2 step 12's turn lane needs. This teaches regeneration to add, drop, and rewrite connections in place.

The dirty-set contract (junctions_are_current)

Discovery found the M2 restriction was not where the issue assumed: add_lane set topology = true but never populated .junctions, so the editor's regen loop iterated an empty list. And DirtySet::topology was doing double duty — "roads/junctions added or removed" (a view concern) and "this command already handled its own junctions, don't regenerate" (a correctness concern). Those coincided only because every topology command so far also built junctions; a lane appearing is topology and needs regeneration, which one flag cannot say.

So DirtySet gains junctions_are_current, set by the commands that build or tear down junction structure (create/delete junction, split_road, delete_road). The editor's regen loop keys off it instead of topology. Default false fails safe: a command that forgets it gets a redundant regeneration (slow, correct), never a stale junction (fast, wrong).

The regeneration restructure

regenerate_junction is rebuilt around a creator. It partitions the freshly planned turns against the existing connection table into matched / new / dropped (keyed by the incoming/outgoing road+contact+lane, so a re-ordered plan is not a changed turn set). Matched turns keep their connecting-road ids; new turns get fresh connecting roads via a factored-out materialize_connection; dropped turns are erased with everything they own (sections, lanes, and any objects/signals anchored to them — erase_road_exact does not cascade).

A TurnSetPolicy parameter keeps the per-frame node-drag preview in-place-only: a preview command is reverted and destroyed every frame, and Arena::erase_exact reserves a slot rather than recycling it, so creating connecting roads there would leak arena slots for the rest of the session. The drag keeps today's behaviour (stale junction + toast); lane edits, which are not previewed, regenerate normally. (The general preview-session leak is filed as a separate issue gating #218.)

insert_lane

Interior lane insert with renumbering + link remap, moved here from #262 — Lane Form and Lane Carve need a lane that starts partway across the carriageway, which add_lane (outermost-only) cannot author. Since odr id is the lateral position, an interior insert renumbers the outer block and remaps everything that named those lanes by id (adjacent-section links, junction lane_links); the inserted lane is unlinked from its neighbours (asam new_lane_appear).

Verification

• 864 ctest (was 854) + 99 pytest green; ASan+UBSan clean on the edit/lane/junction/network suite; soak PASS on seeds 1/2/7 with the new insert_lane op exercised.
• The no-op guard RegenerateJunctionIsByteEqualWhenNothingChanged stays green and was strengthened to also assert a clean revert.
• New tests cover: turn set grows / shrinks (undo resurrects the exact ids), survivors keep their ids, InPlaceOnly still refuses, add_lane names its junctions without claiming they are current, insert_lane renumber / unlinked / neighbour-remap / junction-lane_link-remap / rejections, and the editor path (add_lane on an arm regenerates with no regeneration_skipped toast). The grow and shrink tests were mutation-checked to confirm they bite.

🤖 Generated with Claude Code