atelier-dispatch — write JSONL, dispatch, verify (the lean architecture)

This is atelier rebuilt on the speed-run / local_code_gen substrate: the LLM writes a spec; a deterministic script runs the per-sprint model calls. It exists because the subagent version is dominated by orchestrator cost — each subagent carries 40k of harness context and its report lands back in the orchestrator's context, which then gets re-read every turn (measured: 1.2M cache-read, $1.86 on a 3-sprint task). This architecture avoids both: bare calls ($0.02–0.05 each) and outputs go to disk, not into your context.

Why it's cheap (the two fixes)

1. Bare calls. Each sprint runs claude -p "<prompt>" --model {haiku|sonnet} --bare — --bare strips hooks/plugins/system-prompt, so a Haiku sprint call is ~$0.002–0.05 instead of ~$0.05 + 40k harness. Uses your subscription OAuth — no API key.
2. You never ingest outputs. The script writes files to disk and returns a manifest.json. You read the manifest (tiny), not the generated code. No context compounding, no re-read tax.