feat(cdt): add machine-verifiable acceptance criteria to plan template

[rube-de]

Part of #104

Context

CDT's plan template has a "Testing Strategy" section (plan-workflow.md:128-130) with prose descriptions: "Framework, scenarios, acceptance criteria." But there are no structured, machine-verifiable acceptance criteria that the dev phase can self-check without human oversight.

This is CDT's biggest risk in auto-task mode (no human gate). Sean Moran's research shows unstructured multi-agent systems amplify errors by 17x — the root cause is agents deviating from commitments and holding wrong expectations. Machine-verifiable acceptance criteria serve as the gate that a human reviewer would otherwise provide.

GitHub's Spec Kit and Addy Osmani's 70/30 rule both emphasize: specs with testable assertions, not prose.

Task

Add a structured "Acceptance Criteria" section to the plan template that the architect must fill with machine-verifiable assertions. Also add a "Three-Tier Boundaries" section for scope control.

Acceptance Criteria

• VERIFY: Plan template includes ## Acceptance Criteria section with VERIFY: prefixed items
• VERIFY: Architect prompt instructs: "Write acceptance criteria as testable assertions, not prose"
• VERIFY: Plan template includes ## Boundaries section with green/yellow/red tiers
• VERIFY: Dev-workflow parse step extracts acceptance criteria for verification in Final Verification (Step 7)
• VERIFY: Testing Strategy section references acceptance criteria instead of duplicating them

Implementation Guide

Files to Modify

• plugins/cdt/skills/cdt/references/plan-workflow.md — Add sections to plan template (after line 130), update architect prompt
• plugins/cdt/skills/cdt/references/dev-workflow.md — Update Final Verification (Step 7, lines 235-241) to check acceptance criteria

Approach

1. In plan template (plan-workflow.md, after Testing Strategy section, line 130), add:

   ## Acceptance Criteria
   [Machine-verifiable — each must be testable by an agent without human judgment]
   - [ ] VERIFY: [Function/endpoint/component] returns [expected] given [input]
   - [ ] VERIFY: [Performance target: specific numbers]
   - [ ] VERIFY: Zero regressions on existing test suite
   - [ ] VERIFY: [Additional criteria specific to this task]
   
   ## Boundaries
   - 🟢 Agent proceeds freely: [implementation details, naming, internal structure]
   - 🟡 Pause and confirm with lead: [new dependencies, public API changes, schema changes]
   - 🔴 Never without explicit approval: [deleting tests, modifying auth code, changing prod config]

2. In architect prompt (plan-workflow.md, step 6 area), add:

   Acceptance criteria rule: Write every acceptance criterion as a testable assertion prefixed with "VERIFY:". Each must be verifiable by running a command, checking output, or inspecting code — never subjective prose like "improved performance" or "better UX".
   

3. In dev-workflow.md, Step 7 "Final Verification" (lines 235-241), add:

   5. Check acceptance criteria from plan — verify each VERIFY: item passes
   

4. Update Testing Strategy section in plan template to reference acceptance criteria:

   ## Testing Strategy
   [Framework, test types, QA scenarios]
   [Tests should validate the Acceptance Criteria above — each VERIFY item should map to at least one test]
   

Patterns to Follow

• The PM skill's issue templates use VERIFY: prefix for acceptance criteria — same convention
• The plan template's existing "Validation" section (line 134-135) is the PM verdict — acceptance criteria are different (spec-level, not approval-level)

Scope

In Scope

• Adding Acceptance Criteria and Boundaries sections to plan template
• Updating architect prompt with criteria-writing rules
• Updating dev-workflow final verification to check criteria

Out of Scope

• Automated acceptance criteria enforcement via scripts
• Changing the PM validation flow
• Adding Carmack's functional-vs-quality gate distinction (future enhancement)

Dependencies

• Blocked by: feat: add Temporal durable execution skill plugin #8 (researcher-before-architect) — better research → better criteria
• Blocks: docs: add temporal and doppler to CLAUDE.md #10 (council review) — council reviews the plan including acceptance criteria

Definition of Done

• Implementation complete
• bun scripts/validate-plugins.mjs passes
• PR approved and merged
• Epic checklist updated