Plan batches

Rules: planning/batches/README.md. No time estimates.

1. Split enormous units. For each unit scored 12 or more, split it into 2 or 3 units that each score 11 or less: by item group, by module, the protocol or format change on its own, or a spike or decision part first. Move the plan text verbatim (shared context goes to every part). Write new unit files, delete the old one, fix links and the Depends on lines of units that depended on it. Use one subagent per 2-3 units when there are many.
2. Collect for each unit: priority tier (breaking, user feel, user feature, developer tool, platform), effort level, protocol or format change (y/n), main files, dependencies. A unit that another unit needs inherits that unit's priority.
3. Order, batch by batch, from the units whose dependencies are all in earlier batches:
   • 2 low or med units by priority;
   • 1 high unit by priority;
   • 1 unit with the longest chain of work after it (sum of effort totals of its dependents);
   • fill to 5 by priority.
   • At most one protocol or format change per batch; avoid two units with the same main files.
4. Write: move the unit files into batches/NN/, update each # unit (Batch N) title, regenerate each batches/NN/README.md (levels, protocol change, the table), regenerate the batch list in BATCHES.md (each unit with its level and a one or two sentence description), and fix every link.
5. Check: no dependency in the same or a later batch; python3 tools/check-links.py gives 0 broken. Commit.
6. Report: the first 3 batches in a table, what moved and why, and anything that broke a rule.