Feasibility probe — partial seal, 0 runs (block re-confirmed)

Session 01KWZE3BVCERY4G514G3PDM5ZT re-ran the pre-batch feasibility probe that sealed the prior session (01KWWJ0REMMQ995BQRRZG3K6QS, also 0 runs). All three blockers in the brief's 2026-07-07 Status section persist. Per the brief's explicit "this campaign must not be re-dispatched", no runs were launched and no batch budget was spent on a race the design surface cannot yet stage fairly.

Probe findings (2026-07-07T23:20Z):

1. Design surface (primary blocker). sionna-csi-runner capabilities remain {synthetic-csi, differentiable-ray-tracing, crowd-occlusion, channel-simulator, ble-rssi-synthesis, csi-hardware-impairments} — no stratified-los placement key. The brief's core requirement, "≥10 LOS + ≥10 NLOS links per floor by construction", is still inexpressible. Without it the c-csi-topology-sniffer-count imbalance (28/40 rows wall_count=1, 4/40 LOS) would recur — and per the audit's modal lesson, more seeds never fix a biased design.
2. Infra. gis_stats → PostGIS unavailable (127.0.0.1:5434, password-auth failure). Scene staging via gis export-floor is impossible.
3. Corpus. Floorplan inventory shows resplan with no staged files and missing _crs.txt; ZInD absent. The ≥20 ResPlan+ZInD floor target is unmet, and PostGIS being down blocks import regardless.

Success criteria — all unresolved (0 quantitative claims made):

• Primary (topology d≥0.2 CI): no LOS/NLOS-balanced corpus constructible → no Cohen's d estimable.
• Horse-race honesty: no links.parquet corpus → no area-vs-topology CI comparison.
• Method (Welch + exact permutation + BH-FDR): no data to apply it to.

Unblock path (all three required before next dispatch): (a) a stratified-LOS placement mechanism — first-class runner config key or a host-side wall-count scene-staging step; (b) PostGIS restored on the runner; (c) ≥20 ResPlan+ZInD floors staged. This partial seal retires no hypothesis — the ambiguous topology-vs-area state stands until the fair test is runnable.