arbeitsplan-matrixlisted

# The matrix backend The second way to run a swarm: **one fresh `claude -p` process per candidate** instead of one in-session dispatch. It buys what in-session dispatch structurally cannot. A fresh process has a fresh registry, so a just-edited skill is actually loaded. A cell's `--model` cannot be inherited from the caller — the failure named as *"an omitted model inherits your session's model, which silently defeats this section"* is impossible here. And plugin-present versus plugin-absent arms are expressible at all, which in-session they are not. `cases × models × plugin_states × repeats` at **1 × 1 × 1 × N** *is* the swarm: N independent candidates over one scope. ## Run it, after proving it can One cheap `claude -p` call establishes whether the environment can authenticate at all, then the sweep runs. If the probe fails, `run_matrix.sh` refuses **and prints what the probe got back** — which is the information a blanket refusal could never give you. This replaced an unconditional refusal-when-nested that was imported from another repository's measurement and never tested here. It does not reproduce in this environment, and it was redundant besides: a cell that hits an auth banner is already scored `UNMEASURED` with its reason, and excluded from every denominator. Two honest caveats that survive the change: - A long sweep is still better run from a terminal, where it is not competing with a session for attention and where you can watch it. - `--skip-probe` exist