01KV6EFSAPWE9N8ZF903T64BNH

c-resplan-crowd-geometry-gallery — session 01KV6EFSAPWE9N8ZF903T64BNH

One floor-agnostic through-traffic scenario (east→west symbolic regions, resolved per-floor against the derived walkable polygon) ran across a six-floor ResPlan size ladder (resplan-6090 tiny → resplan-1374 largest) × 2 intensities × 2 seeds = 24 runs, all gate-passed.

C1 (movement surfaces) — MET. Every run emitted trajectory.html, footfall_heatmap.png, speed_vs_time.png and a non-empty trajectory.parquet. The crowd is visible on every geometry.

C2 (kinematic + walkable gate) — MET on all 24. n_outside_walkable == 0 and mean_walk_speed == 1.2 m/s (in the 0.8–1.6 Weidmann band), peak ≤ 1.2, on every floor and both intensities. The symbolic-region resolver lands valid spawn/goal sub-areas inside each floor's walkable polygon with no per-floor tuning — the central generalisation result.

C3 (traversal closes at i1.0) — PARTIAL (5/6 floors). departed == spawned == 12, turned_away == 0 on 6090, 10425, 8266, 10156, 1374. On resplan-16157 only 11/12 cleared (both seeds); one agent takes a ~147 s detour — a near-trap in the 4-bed/4-bath layout (a pinched corridor the largest-part walkable likely narrows). A per-geometry finding, visually diagnosable in 16157's replay — not a model failure.

C4 (geometry→flow signal) — MET (primary), weak (secondary). Mean per-agent traversal time rises with floor size: Spearman(bbox_area, mean_traversal_s) = 0.83 (p = 0.04) across the six floors (6090 5.3 s → 1374 15 s; 16157 35 s, inflated by the trap agent). Under 2× load, full clearance degrades with complexity (8266/10425/10156: 24/24; 6090/1374: 23/24; 16157: 22/24).

Verdict. The floor-agnostic scenario generalises across the ResPlan corpus and traversal cost tracks geometry — support for the Indoor Crowd Modeling generalisation claim. Follow-ons: root-cause the 16157 near-trap (walkable connectivity); reach a congestion regime (see criticism).

Attached runs: 24 (gate 01KV6EM64ER63KVHH6M00CW49H + 23 grid). Per-floor replays at sim://run/<id>/artefact/trajectory.html.

Critic notes (self-review)

• No congestion regime reached. mean_walk_speed is 1.2 m/s on every run, including intensity 2.0 — the crowd is free-flowing throughout. So C4's primary clause (traversal-time vs size, ρ=0.83) holds, but its secondary clause ("congestion onset grows with choke points") is supported only weakly, by the non-departure counts, not by any speed slowdown. To exercise the fundamental-diagram regime these residential floors need far higher agent counts (12–24 agents in a multi-room apartment never reaches ~2 ped/m²). Treat C4 as "size→traversal confirmed; load→slowdown not demonstrated."
• Spearman on n=6 is borderline (p=0.042); the correlation is driven partly by 1374's size and 16157's outlier. Directionally sound, not a tight estimate.
• 16157 outlier is unexplained. The 147 s agent / 11-of-12 clearance is flagged but not root-caused; could be a walkable-derivation artefact (a doorway the wall-buffer nearly seals) rather than genuine geometry. Needs a look at that floor's largest-part walkable before the result is cited.
• Single arrival-window + 2 seeds. Not a robustness sweep; the per-floor means are point estimates.