fix(simulation): coastal-cell routing relaxation — roads return to water-heavy bodies #215
Assignees
jpmschweitzer (Jeroen Schweitzer)
Clear assignees
jpmschweitzer (Jeroen Schweitzer)
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: jpmschweitzer/settled-reach#215
Reference in New Issue
Block a user
Summary
Batch-7 server track: T-1116. Several real bodies (GJ251c, GJ380c) generated settlements but ZERO road edges — every A* pair dropped because a coastal settlement's land pixel can sit inside a routing cell that is water-majority after downsampling (a cell folds up to 64 native pixels), and astar() hard-returned None for any pair touching an IMPASSABLE anchor.
The fix (single file, road_graph.rs): relax only the start/goal anchor lookup.
nearest_passable_cell(ring BFS, deterministic row-major tie-break, bounded at COASTAL_ANCHOR_MAX_RING=3) finds a surrogate anchor; the route pays COASTAL_ACCESS_COST_PER_RING per ring — a short, honestly-priced access road, never a free water crossing. IMPASSABLE semantics untouched everywhere else (D-210 transit costs, open ocean). The placement-nudge alternative was rejected: it would move Layer-3 state D-211 promises is seed-derived.Boundary semantics pinned by test: exactly-half-water cells stay passable (strict-majority rule); a settlement with no passable cell within the search ring degrades to an isolated 0-edge node — never a panic, never a fabricated route.
Test plan
Review: road-routing-coastal-fix -> main (type: code)
Reviewers: Hoshe (QA) + Tyre (architecture/governance). Both REQUEST_CHANGES.
Hoshe (QA): REQUEST_CHANGES
length_cellsis a pure HOP COUNT (a grassland step contributes 1 regardless of its 200-unit cost), but the surcharge is computed in cost units thendiv_ceil(MIN_CELL_COST=50)-folded in — yielding 4 hops per ring, not the documented "one extra grassland hop." Worst case (both anchors ring 3) adds +24 length_cells — double WAYPOINT_THRESHOLD_CELLS — spuriously tripping (or elsewhere masking) the add_waypoints gate for physically short edges. Fix: derive the surcharge directly in hop units (or a separate field the waypoint gate doesn't read) and correct the comment's stated rationale.min_edges = 1while the commit message and the test's own module comment claim "GJ251c 0→2" — a regression to exactly 1 edge would pass silently. Tighten to the documented 2.Hoshe's confirmations on record:
nearest_passable_cellgenuinely deterministic (plain Vec, fixed push order, no hash iteration), correctly bounded (ring 0 = self;1..=MAX_RING, no off-by-one) and grid-edge-safe (rows clamp, columns wrap); IMPASSABLE untouched for mid-path transit (main-loop guard diffed against main); the drawn geometry force-anchors both ends at the exact settlement position — no visible road-to-marker disconnect risk for the eyeball; the determinism test compares full structural graph equality (PartialEq over edges/costs/paths/nodes), not counts.Tyre (architecture/governance): REQUEST_CHANGES
!ocean_mask"always land" guarantee lives in features.rs::extract_attractors (D-209, lines 544/570) — not attractor_matching.rs (D-211), which does no ocean filtering. Fix the pointer, and add a word on the D-211 phase-4 synthetic-overflow path (PlainCenter) — verify it also guarantees land, since synthetic attractors are generated positionally rather than extracted.COASTAL_ACCESS_COST_PER_RING = ORTHO * 100re-hardcodes the grassland baseline100(cost-field initializer) as a bare literal — if the baseline changes, the surcharge silently decouples from its own rationale. Extract a shared named constant. (Interacts with Hoshe #1 — resolve together.)Tyre's confirmations on record: the fix-choice adjudication is accurate (D-211's positions genuinely seed-derived; layer analysis right); D-210 else-branch transit semantics byte-identical to main; determinism pure-integer with no cross-platform hazard; believability provably inert to roads (zero road_graph references in believability.rs; the layer runs but no metric reads it); scope containment decisive —
length_cellsis dropped at the Atlas wire boundary (layer_proxy.rs:1796/1800) and maintenance_authority uses straight-line euclid, so the surcharge never reaches the client.Verdict: CHANGES REQUESTED
All five findings dispatched to the authoring agent; fix round → gate → live road-overlay eyeball (GJ251c, GJ380c) → merge.
Fix round — all 5 findings addressed in
de8bcf4ebCOASTAL_ACCESS_SURCHARGE_HOPS_PER_RING = 1, added directly tolength_cellswith no unit conversion (the old cost-unit constant div_ceil'd to 4 hops/ring; worst case was +24 — double the waypoint threshold — now +6). A formula-pinning test asserts both the arithmetic (in isolation via the start==goal short-circuit) and the constant's value, so regression requires a visible deliberate change. Edge counts on both repro bodies verified unchanged (reachability was never affected — the bug only inflated reported length).features.rs::extract_attractors(D-209). The instructed check of the D-211 Phase-4 synthetic-overflow path found a REAL gap:synthetic_attractortakes no terrain argument and never checksocean_mask— a synthetic-overflow settlement can land in open water. Documented in the code comment and the D-210 amendment, not papered over; T-1206 filed (the routing relaxation degrades that case gracefully, but the placement guarantee is Layer 3's gap to fix).pql decisions validateok; sync clean, 0 broken refs.100in the file is the pre-existing cost-vector pre-fill (immediately overwritten), untouched as out of scope.No retractions — all five findings accurate. road_graph suite 25/25; full cargo test 0 failures (1977 lib + all golden harnesses). Gate green on push. Live before/after road-overlay eyeball (GJ251c, GJ380c) in progress — merge follows it.
Pull request closed