Session synthesis — c-csi-cross-geometry-scaleout (Phase 1 pilot)

Outcome: BLOCKED on a hard prerequisite. Zero runs produced artefacts; no success criterion is resolved.

What happened

A 3-floor validation batch (resplan-12439-multiroom, resplan-1374, resplan-147440; seed 0; occupancy grid {0,2,4,6,8} @ 2.4 GHz) was fanned out before committing the full 10-floor × 3-seed × 5-level (150-run) Phase 1 pilot. All three subagents failed identically at scene staging: PostGIS authentication failed for user monad on port 5434. exp-csi-static exports each floor's scene from PostGIS via gis export-floor before ray-tracing, so an unreachable/mis-authenticated DB blocks every run. This is the external-dependency risk the brief pre-declared (a topology-sibling session hit connection-refused on 2026-07-06; today it is an auth failure on :5434, not the :5435 the brief named).

Criteria status

• Primary (LOFO inflation-ratio CI): UNRESOLVED — no CSI, no MAE, no bootstrap. (Already pre-declared under-powered at n=10 even on success; now simply un-run.)
• Variance split (σ²_floor vs σ²_seed): UNRESOLVED — needs ≥2 floors × 3 seeds of artefacts; zero produced.
• Comparator honesty (LOFO denominator): UNRESOLVED — the honest leave-one-occupancy-out baseline cannot be computed without runs.

Machinery-validation finding (a Phase 1 goal, achieved)

The launch path was exercised far enough to confirm a schema constraint: StaticParamsSchema is extra_forbidden and rejects a nested grid: dict with array-valued n_agents/radio.carrier_freq_hz. A single run takes flat scalar params; the occupancy/frequency grid in the campaign sim_params.grid block must be expanded at the sweep/launcher layer (one run per cell), not passed inside a single run's params.

Next action (blocking, for re-dispatch)

1. Fix PostGIS reachability + credentials for user monad from the runner env (note the port drift: brief says :5435, runner hit :5434 with a password-auth failure). Verify with gis export-floor on one floor before dispatch.
2. Fold the occupancy grid into the sweep layer (sim_sweep_launch) or emit one run per (floor, seed, n_agents) cell — do not pass array-valued grid in per-run params.
3. Re-dispatch Phase 1 once (1) is green; it remains a pre-declared under-powered pilot for the primary criterion, resolving only the variance split + machinery.

Figures

None rendered — zero artefacts exist to plot. The mandatory csi_cross_geometry / seed_trace figures are deferred to the first session that produces runs, not omitted.

Budget: CPU essentially untouched (no ray-tracing ran); session sealed early to avoid spending against a dead prerequisite.