feat: integrate Attractor DOT pipeline patterns into disciplined-* skills

[AlexMikhalev]

Summary

Integrate declarative pipeline patterns from StrongDM's Attractor into the disciplined-* skill family. Each skill gets a companion pipeline.dot file defining its workflow as a Graphviz directed graph with mechanical gate enforcement, checkpoint/resume, and context fidelity management.

Based on: https://github.com/strongdm/attractor (788 stars, Apache-2.0)

What Changes

Current	Target
Workflows defined as SKILL.md prose	DOT graph companion files (visual, diffable)
Gate passage is advisory (agent recommends)	Mechanical enforcement (orchestrator refuses to proceed)
No crash recovery	Checkpoint after each node, resume from last completed
Context all-or-nothing between stages	Fidelity modes per edge (full/summary/truncate)

Implementation Phases

Phase 1: DOT Pipeline Definitions (~2h)

Create pipeline.dot for each disciplined-* skill:

1. disciplined-implementation
2. disciplined-verification
3. disciplined-research
4. disciplined-design
5. disciplined-specification
6. disciplined-validation

Each pipeline: named nodes with typed handlers, conditional edges for pass/fail routing, goal_gate=true nodes for quality checkpoints.

Phase 2: Goal Gate Enforcement (~1.5h)

• Parse goal_gate=true from DOT
• Orchestrator verifies all gates pass before stage transition
• Failed gates route to retry target (loop, not skip)

Phase 3: Checkpoint/Resume (~2h)

• Save state after each pipeline node
• Resume from checkpoint on context compaction or crash
• Audit trail of completed steps

Phase 4: Context Fidelity Modes (~1h)

• full within implementation loops
• summary:high for inter-stage transitions (design -> implementation)
• truncate for verification -> validation (only pass/fail matters)

Total estimated effort: ~6.5h

Acceptance Criteria

• All 6 disciplined-* skills have companion pipeline.dot files
• DOT files render as valid Graphviz diagrams (dot -Tpng)
• Goal gates are mechanically enforced (not advisory)
• At least one pipeline can checkpoint and resume after interruption
• Context size at inter-stage boundaries is measurably reduced

Not Adopted (Essentialism)

• Full Attractor engine (we orchestrate via skills, not a custom engine)
• Unified LLM Client (terraphim-llm-proxy already handles this)
• Coding Agent Loop (we use Claude Code itself)
• Model stylesheet (model selection is per-skill, not per-node)

Related

• feat: implement .well-known/skills/ discovery endpoint (Cloudflare RFC) #57 (.well-known/skills/ discovery -- DOT pipelines could be discoverable)
• feat: semport skill -- semantically-aware cross-tool and cross-language code porting #59 (semport skill -- same StrongDM Factory family, semport pipelines could be DOT-defined)
• Compound engineering adoption plan P1 (multi-agent review maps to parallel handler + fan-in)

Plan

See: plans/attractor-disciplined-development-integration.md in cto-executive-system