ctde_v0.run_conn_shootout¶
Connectivity-mechanism shootout — the 90/90 assault at 32²/10.
Throws every connectivity mechanism the trainer has at the coverage<->connectivity wall, to find the one that holds connectivity while maximising coverage. Directly tests the deep-research verdict: a LEARNED Lagrangian connectivity constraint (lagrangian / pid_lagrangian — RCPO / PID-RCPO) should beat the brittle fixed barrier-penalty and the soft penalty; the hard action mask is the other strong candidate (our prior "guardrail beats soft by ~20pts").
MECHANISM (4): maskhard (action_mask, hard guardrail) · soft (soft_lambda penalty) · lag (lagrangian = learned dual) · pidlag (pid_lagrangian = PID dual, Stooke 2020). ARM (2): role (explorer/relay) · base (homogeneous). Fixed: 32²/10 · comm_r 5 · collision-mask on · frontier-attn · 100 steps · up-weighted coverage (--w-coverage 3) · conn-signal local_edge_margin (per-agent degree ≥ 1).
Per seed: 4 × 2 = 8 runs. Run-dir
PYTHONPATH=.:../../../FiedlerValueEstimation ~/ZymeraLab/.venv/bin/python ctde_v0/run_conn_shootout.py --out runs/connshoot --seeds 1 --iters 2000 --jobs 3