test-plugin-harnessinglisted

# Testing a phase of the plugin's harness `launch-llms-benchmark` asks *how good is this model's report*. This asks a different question: **how much work does the package still make the model compose before it can do anything** — and it answers it for one named phase at a time. Do not run the benchmark protocol for this; it grades findings, costs a full run, and its row is not what moves when a skill stops making the model write a script. The rules a change here must satisfy are `AGENTS.md`'s **Plugin harnessing** section. This skill is how you prove one landed. ## What you need before starting - **The phase**, named by the caller: `preflight`, `drive`, `observation`, or `whole`. Ask if it is not named — measuring the wrong phase wastes the run. - **The mission**, the same on every sample, and by preference an observation in **drive mode on the local stack**: the run generates its own traffic, so every window holds the same requests and the samples differ in how they worked, not in what there was to find. A post-hoc window holds whatever landed that minute (a late burst, a real error — measured 2026-09-09: one 502 gave three samples an extra finding and 34 to 40 queries against 24 to 30): measure post-hoc only when post-hoc is what changes, with one scripted burst per sample and the stack's facts re-read per window; a remote stack only when the change is that stack's, never by default. - **The baseline**, which is the published row in `.llms-benchm