01KW022VJ8HQ5ACTE73XR6Q109

ResPlan crowd-flow gallery — full corpus (10 floors)

One unmodified scenario (symbolic east→west through-traffic, from_floor walkable derivation, agent_clearance_cm: 2.0, 12 agents, seed 0) was swept across all 10 ResPlan floors × intensity {1.0, 2.0} — 20 runs (sweep 01KW01RGSNM586FYECQ8DSVMC4).

C1 — met (20/20). Every run carries trajectory.html, footfall_heatmap.png, speed_vs_time.png + non-empty trajectory.parquet. The gallery contact-sheet renders all 10 footfall surfaces.

C2 — met (20/20). n_outside_walkable == 0 everywhere; mean_walk_speed 0.81–1.18 m/s (in the Weidmann band); peak is pinned at the 1.20 m/s cap (a cap, not an emergent peak).

C3 — not met. 9/10 floors close at intensity 1.0; resplan-7421 strands 1/12 agents (departed 11, turned_away 0) at a 0.68 m² pinch near (14.9, 8.9). This is floor-specific — resplan-6090 departs 12/12 despite an explicit impassable-risk QC flag, so the failure is not endemic — but note the harness has no stranded-agent counter, so the loss passes turned_away==0 through an accounting gap rather than correct bookkeeping. (At intensity 2.0, 6090, 7421 and 12439 each strand 1–2 agents under load.)

C4 — unanswerable as pre-registered. mean_traversal_s does not exist in the walk-notebook v1 schema, so the Spearman test cannot be run. A speed-degradation proxy (i=2.0/i=1.0) is directionally consistent (constrained floors degrade, open floors barely), but its strongest point (7421, 0.68×) is confounded by agent stranding and is not reported as clean congestion evidence.

Scope. Single seed; zero-occupancy through-traffic only. This supports kinematic portability of the scenario layer across the ResPlan size ladder (32–357 m²), not a general crowd-modelling claim. Follow-ons: add a mean_traversal_s / per-agent departure-timestamp field and a stranded-agent counter to the runner; review 7421 + 6090 geometry.

Critic (severity: medium, agrees with verdict)

Claim audit. C1/C2 solidly supported across sampled runs. C3 fails as stated. C4 unanswerable (schema lacks mean_traversal_s).

Weak claims corrected in the final synthesis: (1) C3's "geometry finding, not model limitation" let the harness off the hook — there is no stranded counter, so the lost agent passes turned_away==0 by accounting gap; now flagged. (2) The C4 7421 0.68× proxy is the most contaminated point (7421 i=2.0 strands 2/24, so low speed mixes stuck-agent drag with congestion) — no longer the headline. (3) "peak<=1.20" is a pinned cap, not an emergent finding — clarified. (4) "Indoor Crowd Modeling generalisation" softened to scenario-layer kinematic portability.

Scope: single seed (every cross-floor delta within plausible single-seed noise, no CI); congestion onset not cleanly readable because the speed proxy is confounded by stranding on exactly the constrained floors of interest.