Condition-switch fusion — Session 01KWT73W (partial, budget-sealed)

Outcome: scenario authoring delivered and validated; the compute-heavy run grid + reduction are deferred to a CI follow-on. This session banked the campaign's highest-risk step (expected-interpretation #4: a failed burstiness manipulation would invalidate every downstream fusion number) and stopped short of the 6 ray-traced coupled runs, which the supervisor's LLM headroom could not fund without risking an over-budget, orphaned close.

Authoring (de-risked). Both occupancy-dynamics arms were authored as walk-notebook from_floor kind: groups casts (no agendas/fsms, per brief) and pass sim_scenario_validate (schema, duration_units: 3600):

• smooth — one group, n_agents: 18, single wide arrival_window_units: [0, 300] → roughly constant arrival rate.
• bursty — three groups of 6, narrow staggered windows [0,20] / [120,140] / [280,300] with quiet plateaus → 2–3 sharp waves, same total headcount (18) and comparable occupancy range.

Each merges over sim_params.exp-csi-crowd.base (ble.enabled, radio 2.4 GHz/20 MHz/52 sc/depth 3, body_loss_db 5.0, trajectory_frame_stride 24, Doppler NOT enabled) with seed ∈ {0,1,2} → the 6-run grid.

Criterion status. C1–C4 unresolved (budget): no runs launched, so the empirical burstiness check (≥0.15 busiest-20% |Δocc| separation), the fusion_condition_switch{,_sensitivity}.parquet MAE table (BLE-held / CSI-adapted / clock tau=16 s / condition-switch d0,kappa from calibration window / oracle), the headline bursty crossover, and the tau∈{8,16,32}×kappa×{0.5,1,2} tuning-honesty sweep are all pending. C5 upheld: this synthesis claims nothing beyond a validated authoring recipe; the crowd is self-authored and the BLE counter modelled — no ble-periodic-calibration / recalibration-trigger-from-drift strength upgrade is asserted.

Follow-on. Re-run via /campaign-systematic c-csi-ble-condition-switch (CI runner, project key) which can afford the 6 RT runs within 1.0 CPU-h + the reduction. Launch each arm with the validated scenario block above; if the notebook container cannot read run parquet from Hetzner S3, apply the brief's standalone-reduction escape hatch. No critic/patcher pass was run — a partial synthesis with four explicitly-unresolved criteria is low-severity by construction.