feat: Add --checkpoint-between-waves flag to devplan_generate_executor

[mmorris35]

Summary

Add an optional flag to devplan_generate_executor that injects /clear + re-read instructions between waves in the generated executor agent.

Motivation

Based on Ashley Ha's "subtractive context management" methodology (UC Berkeley ML, prev. Rivian):

• Never exceed 60% of context window
• /clear between phases prevents instruction drift
• Fresh context = Claude actually reads instructions with full attention

Thread with 5.2K views validating this approach: https://www.threads.net/@jatinw21 (Feb 15, 2026)

Proposed Behavior

devplan_generate_executor --checkpoint-between-waves

Generates executor agent with instructions like:

## Wave 1: Core Infrastructure
[subtasks...]

---
⚠️ CHECKPOINT: Before starting Wave 2:
1. Run `/clear`
2. Re-read DEVELOPMENT_PLAN.md sections for Wave 2
3. Continue with fresh context
---

## Wave 2: Feature Implementation
[subtasks...]

Benefits

• Prevents context bloat on large projects (5+ waves)
• Keeps Haiku focused on current wave only
• Aligns with validated methodology from ML practitioners
• Optional flag — doesn't change default behavior

Scope

Small — mostly template changes in the executor generator.

Related

• Subtractive context management pattern
• 60% context window rule