EXP-S1 Geometry A fundamental-diagram gate — session 01KTNHCDZG

This session re-ran the FD gate (CFSM defaults; n_agents ∈ {2,4,8,16,32,48}; seeds 0/1/2; 180 s, dt=0.05, transient_frac=0.17) after fixing three harness defects that had silently blocked the campaign — and that the prior CI attempts never surfaced (they declined at turn 0 on budget grounds):

1. pedpy-analyser crashed on len(PosixPath) before writing metrics/figures (every run failed the gate).
2. compute_voronoi_speed was called with the wrong PedPy-1.5 signature → silent fallback to a naive speed estimator.
3. The single central-waypoint scenario parked agents on arrival, collapsing post-transient Voronoi speed to ~0.04 m/s at every density (degenerate FD). Replaced with a cyclic closed-loop journey so a fixed crowd circulates.

Criterion 1 (four regimes) — NOT met, by geometry. Room-average Voronoi density spans only 0.02–0.50 ped/m² (free-flow + light-congestion). Dense/jam (ρ≥2) require ~400 agents or a constricted measurement area on this 96 m² room — the brief's documented out-of-scope limit. 2 of 4 regimes populated.

Criterion 2 (≥75% in-band) — MET: 14/16 = 88%. (ρ,v) pairs fall inside the Liddle-2022 / Adrian-2024 band. The two outliers are n=32 (v≈0.78 at room-average ρ=0.33, nominally free-flow): the closed loop concentrates agents, so local crowding — and the speed reduction it drives — onsets at a lower room-average density than the room-keyed band predicts. This is a measurement-area effect, not a model failure.

Criterion 3 (monotonic v↓, per-seed Spearman ≤ −0.6) — MET decisively. ρ(n_agents, v) = −1.000 for every seed; mean speed falls 1.17→0.68 m/s as crowding rises, with negligible seed variance (σ ≤ 0.018 m/s).

Verdict (brief's scenario 2): the EXP-S1 CFSM-default stack reproduces the Adrian–Liddle fundamental diagram across the density regimes reachable on Geometry A. Phase B is gate-cleared for free-flow + light-congestion, with two flagged follow-ons: (a) a bottleneck / larger geometry to reach dense/jam, and (b) a local-density measurement area to resolve the room-average mismatch at high n.

Caveats. JuPedSim 1.4.2 seeds perturb only spawn jitter, so the three seeds are near-replicas (tight σ) rather than strongly independent draws — the per-seed Spearman is robust but the seed axis is weak. 16 of 18 planned runs were captured (two seed-2 points, n=16 and n=32, did not register). The three harness fixes (analyser ×2, scenario) are uncommitted working-tree changes pending review.