[idea] Context diet: Stages 0–4 should start at ≤25k toolkit text (runbook spine, run-don't-read, per-stage routing) #166

The idea: cut the default context load of early-stage sessions. Measured with chars/4:

• conversion-runbook.md is ~23k and is read in full every session.
• interview-protocol.md is ~5k.
• The stage "-" baselines are ~7k.
• Script rows are ~25k when an agent reads a script instead of running it.

That comes to about 40–70k before the task starts, and it is re-read on every call. Proposals:

1. Split the runbook into a ~4–5k spine plus per-stage sections. This is the biggest saving, ~18k per session.
2. Change the bin/*.sh routing rows to "run with --help", so agents run scripts instead of reading them.
3. Render stub routing per agent and per stage.
4. Pass agents the slice they work on, not the whole spec.
5. Add a "context at start: Nk" line to each Live Checklist.

Why this belongs in the toolkit: it applies to every entry mode, since every session reads the runbook.

Field evidence: ProcureFlow cook-off. Details in contrib/inbox/2026-09-29-context-diet-early-stages.md (#165).

Done = a Stage 0–4 session starts at ≤25k toolkit text, measured from the first call of a transcript, with no gate regressions (render-routing.sh --check, stage fixtures).

Rough size: a cluster (runbook, routing TSV + renderer, stubs).