Lab run dispatcher

Large-profile ensembles never run in the cloud container (4 cores). They run on the user's 24-core Windows machine. Your job is to decide where a run belongs, then hand over exact lines.

1. Estimate cost

CPU-seconds = seeds x (fork days + branch-days) x 30
branch-days = (listed branches + 1) x days — the control is always run in addition to the listed branches (runLab runs the control and every branch from the same save). If --branches is omitted, lab_seed.ts defaults to DEFAULT_BRANCHES (nudge:1, nudge:2, no-wolf-memory, random-choice, no-wood, no-food, no-clay): 7 + control = 8 branch-days.
Wall time = CPU-seconds / cores. First run nproc: 24 means you are on the user's machine and the lines are for here; 4 means the cloud container and the run is handed over instead.

• One --dynamics per run: to compare rich against authored, run twice with two --out directories. Never mix dynamics in one ensemble.
• Use 30 CPU-s per simulated day (measured 28.1 and 29.8 with all 24 cores busy). Never use the idle-core 15; it halves every estimate. Past check: plans ensemble predicted 20 min, took 28.
• Over about 3 minutes of wall time here, or many Large seeds: give lines for the user's machine. Unit tests, tsc, one-seed smoke runs and short checks run here.

2. Build the command block

Rules for Windows PowerShell 5.1: one command per code block, no &&, no bash syntax.

Always start with the sync lines (each its own block): cd "C:\dev\Claude Experiment\living-world", git pull, and npm ci only the first time or if package.json changed. Remind that the code must be pushed first.

Then the run, from this template:

node scripts/lab.mjs --out <out> --seeds <N> --fork <F> --days <D> --dynamics <authored|rich|rich-unversioned> --branches <list>

and the report line:

npx vite-node scripts/lab_report.ts -- <out>

Facts to respect:

• --seeds max is 32 (12 named seeds + nov-1..nov-20). --dynamics must be one of rich, authored, rich-unversioned.
• Branch names come from AGENTS.md: nudge:N, remove:<types>:<radius> (aliases no-wood, no-food, no-clay), random-choice, no-wolf-memory, rich, rich-stakes-only, rich-mood-only, rich-no-plans, rich-no-ask, rich-no-learning, rich-no-trades, rich-no-stock. Reject anything else.
• Always include at least nudge:1 (better nudge:1,nudge:2): compare a branch against the mean of control and nudges, never one control alone.
• Use a fresh --out directory after any source revision change; the resume guard checks rules version only, not source provenance.
• Novel-need runs use node scripts/novelneed_all.mjs --out <out> --seeds N --fork F --days D instead.

3. Tell the user

Always end with:

1. Estimated wall time on 24 cores, with the arithmetic shown, labelled as an estimate.
2. Paste back only <out>\report.txt, never progress lines.
3. Re-running the same command resumes unfinished seeds (finished seeds are skipped).
4. For anything over about 10 minutes, point them to the windows-detached-launch skill so the run survives the session.