plan-idea-genlisted

# plan-idea-gen — wide-then-narrow ideation anchored to a plan.md Generate POOL ideas (default 100) across forced-diversity lenses, then filter to KEEP finalists (default 10) that survive clustering, pairwise ranking, and an adversarial kill pass. Every run is anchored to exactly one plan.md: it grounds generation, hosts the ledger, decides the bead workspace, and receives accepted ideas. ## Execution model: ONE context, no agents — MANDATORY The entire pipeline runs **inline in the invoking context as sequential reasoning passes**. Never use the Workflow tool. Never spawn Agent subagents — not for generation, not for judging, not even an Explore scout. This is a hard rule, not a default (`allowed-tools` above deliberately omits Agent and Workflow). Why: every subagent pays ~65–70k tokens of fixed context overhead (system prompt + tool schemas) before doing any work, and subagents share no cache. A fanned-out run of this pipeline (~50 micro-agents) costs 3–4M token-weight; the identical logic run inline costs little more than its own output. The judging stages are exactly the kind of small decisions that do not justify a fresh 67k context each. Mitigate the loss of true generator blindness by ORDER: generate the full pool (Step 2) before doing any judging (Steps 3–7), never revise earlier ideas while generating later ones, and re-read CONSTRAINTS immediately before the kill pass. Lenses still buy coverage even without process isolation. Durability: after generation and