Add prompt adherence check to pipeline

[neoneye]

Summary

• Adds a Prompt Adherence step that checks how faithfully the final plan follows the original user prompt
• Two-phase LLM approach: Phase 1 extracts directives (constraints, stated facts, requirements, banned words, intent), Phase 2 scores each against the final plan
• Each directive scored with adherence_5 (1-5 Likert) and categorized as Fully honored, Partially honored, Softened, Ignored, Contradicted, or Unsolicited caveat
• Produces a markdown report with summary table (sorted by ID), overall adherence percentage, and detailed Issues section for all non-perfect items
• Last section in the final HTML report
• Splits plan.txt into plan_raw.json (raw user prompt + date as JSON) + SetupTask (template). Prompt adherence reads the raw user prompt directly, not the templated version.

Motivation

PlanExe's pipeline has a "normalization bias" — each step nudges the plan toward what a reasonable project should look like. Over ~70 nodes the cumulative drift is significant. User-stated facts get overridden, requirements get softened, and the plan adds unsolicited feasibility studies. This step surfaces that drift so the user can see it at a glance.

Test plan

• 7 unit tests for Pydantic models, markdown generation, and score calculation
• 330 total tests pass (no regressions)
• extract_dag picks up the new node (71 nodes)
• Manual test with real pipeline runs

🤖 Generated with Claude Code