Session 01KXBWYQH23670XAGB9RQJP9Z8 — c-sim2real-count-transfer (Phase 2a corpus generation)

Executive summary

This session unblocked the dependency chain that sealed the prior session (01KXBTWW2CM855Z7CW1W5RD4M5) as BLOCKED, and validated the corpus-generation recipe on a single probe cell — but supervisor token/USD budget was exhausted before the full 24-cell class-balanced corpus could be generated, so the three corpus-verification criteria remain unresolved (partial).

What was established (preflight dependency-chain gate)

Three concrete blockers were found and resolved via cheap probes (~1 CPU-min each), exactly the leverage the Step-2.5 gate exists for:

1. sim_sweep_launch does not merge sim_params.base. It launches the Cartesian product over the simulator default config, whose floor is a_seminar → SceneStageError: floor not found: 'a_seminar'. Fix: fold every sim_params.exp-csi-crowd.base constant into the grid as a single-value axis (floor resplan-12439-floor-0, geometry=from_floor, the full radio.* / impairment_profile.* / scenario.* / body.* blocks, doppler.mode=cfr, dt=0.05, trajectory_frame_stride=10).
2. where=remote-vm is unusable on this host — RuntimeError: remote-vm execution requires MONAD_SIM_REMOTE_HOST. The env var is unset. The sionna-csi-runner family is not on GHCR, so where=local is the correct locus (Step 4) and works.
3. Artefact contract verified. With (1)+(2) fixed, probe cell exp-csi-crowd (n_agents=3, seed=0, 5 GHz, cfr, IP-101 impairment profile) ran to completion (run 01KXBX1WN7WHKRQ04WKJPYA4M6) and emitted the full contract: csi.hdf5 (79.5 KB, clean oracle), csi_impaired.hdf5 (672.6 KB — ~8× larger, impairment model demonstrably active), links.parquet (14.4 KB), plus sionna-csi-runner.metrics.json, staged scene, trajectory. No staging error.

Per-criterion status

• C1 (artefact contract): verified for the recipe on one cell — both hdf5 files + links.parquet land, no error. NOT resolved as stated ("every cell") because only 1 of 24 cells ran. Down-scoped to a recipe-existence result.
• C2 (impairment non-trivial + occupancy-preserving): UNRESOLVED. The impaired file being ~8× the clean size is suggestive but not a tested mean_atten_db_per_person_impaired / variance_monotonicity comparison across n_agents. Needs the full grid + a reduction.
• C3 (0–5 × 4-seed class-balanced coverage): UNRESOLVED. Only n_agents=3, seed=0 was run.

Exact working grid for the next session (hand-off)

sim_sweep_launch(sim_id="exp-csi-crowd", where="local", session_id=<new>, grid={...}) with the multi-value axes scenario.n_agents:[0,1,2,3,4,5], seed:[0,1,2,3], radio.carrier_freq_hz:[5.0e9] and every base constant folded as a single-value axis (as in this session's successful probe3 launch). 24 cells; on local each cell ≈ the probe's ~1 CPU-min. Then verify with sim_reduction_run (multi_seed_ci on the impairment metrics; mann_kendall on variance_monotonicity over n_agents) and render the occupancy-vs-CSI figure. Phase 2b transfer test (csi_sim2real_count_transfer.py) consumes the completed corpus, as the brief specifies.

Real-life meaning

The synthetic-CSI corpus for the sim-to-real crowd-count transfer experiment is now reproducibly generatable — the two silent blockers (base-folding, execution locus) that would otherwise fail every cell are documented with the exact fix. A fresh session inheriting this recipe should complete the 24-cell corpus and all three criteria cheaply and without rediscovery.

Caveats

Single supervisor session under hard budget; corpus is 1/24 complete. All quantitative occupancy/impairment claims (C2) are deferred to the reduction over the full corpus — none are asserted here.