c-exps1-fundamental-diagram — Session 01KS7ZS50AQTVF9RTGH101FQNR

Outcome: framework-state report. The FD physics verdict the brief asked for is deferred because the platform exposed three blocking gaps mid-session, two of which were fixed live and the third of which requires a follow-on PR before the (ρ, v) band can be evaluated.

What ran

37 runs attached, 0.023 / 2.5 CPU-h used. The corpus splits into three cohorts that the analysis must filter on params.phase:

Only the 18 fd-gate-v2 runs are usable for any (ρ, v) reasoning. The other 19 are the audit trail of getting there.

Per-success-criterion verdict

1. All four Adrian-2024 regimes populated — not resolved. n_agents ∈ {2, 4, 8, 16, 32, 48} on Geometry A (96 m² walkable) reaches a global ρ of 0.50 ped/m² at G6, brushing the free-flow / light-congestion boundary; Voronoi-cell peaks reach ~0.625 ped/m² at G4 (from pedpy.metrics.json max_density). Reaching dense flow (ρ ≥ 2) and jam (ρ ≥ 4) needs either a constricted measurement area or ≥200 agents — out of the brief's scope.
2. ≥75% of (ρ, v) pairs in the published band — unresolvable. pedpy-analyser's entrypoint emits no scalar mean-speed; only max_density and mean_residual. The (ρ, v) coordinate the criterion compares against the band cannot be computed from the current artefact schema. This is the load-bearing gap.
3. Per-seed Spearman ρ between n_agents and v — unresolvable for the same reason. Even with speed in hand, the v1 cohort was seed-deterministic (jupedsim-runner only set PYTHONHASHSEED, which JuPedSim's CFSM doesn't read), so per-seed variance was an artefact of measurement noise rather than the simulation itself. The v2 cohort fixes seed-injection but still hits gap (2).

Framework deliverables that did land

The session is best read as a framework stress-test that produced three concrete artefacts:

1. Phase D vault-backed simulator resolver — monad_knowledge/sim/runtime/dispatch.py:resolve_simulator_chain now reads _knowledge/simulators/<slug>.md frontmatter (image + last_seen_digest) as the canonical source; MONAD_SIM_SIMULATORS env is per-slug fallback. The previously-empty last_seen_digest: field on both simulator notes is populated; four new tests in tests/sim/test_cli_resolvers.py cover the vault branch + vault-wins-over-env precedence.
2. jupedsim-runner entrypoint bug fix — the v1 image's start-position generator cycled back to starts[0] for n_agents > 16, generating duplicate spawn coordinates that JuPedSim rejected (Wave 3 failure). The v2 entrypoint uses a deterministic grid (0.6 m pitch on a 6×10 m inset rectangle, capacity ~150) with seeded ±0.09 m RNG jitter so seeds actually produce distinct trajectories. v2 digest sha256:f678c5.
3. MCP env propagation lesson — the MCP server inherits env at start, so per-host runtime config (MONAD_SIM_NETWORK, MONAD_SIM_OTLP_ENDPOINT) belongs in .mcp.json's env block, not in a shell export. Both are added; MONAD_SIM_USE_TAG is no longer required because the vault resolver returns digests that match local RepoDigests after the canonical retag.

Follow-ups (filed as separate IPs)

• IP-A: extend pedpy-analyser to emit mean_speed, mean_voronoi_density, plus per-time-bin Voronoi summaries. Unblocks criterion 2 + 3 on a future session.
• IP-B: --image-source=build|registry flag with sensible per-mode defaults — local = build, remote = pull digest. Crystallises the curious-vs-systematic routing the prototyping is converging toward.
• IP-C: cost-routing convention — /campaign smoke-tests on the laptop; campaigns wider than ~6 sim-runners dispatch via sim-campaign.yml. The supervisor LLM lives on the cloud worker, not the subscription.

Thesis-chain implication

None for chapter 5 yet. This session does not support or refute the L4½ Adrian-Liddle band claim. It does support a meta-claim about the IP-078/081 framework: the platform reaches the FD-gate question but does not yet answer it; one more analyser-side PR closes the gap.