v0.7.2 -- Parallel agents, half the wait

Parallel Agent Orchestration

The build phase was the longest in the pipeline -- a 6-screen project spawned 8 sequential agents for ~35 minutes of wall-clock time. Now the orchestrator plans work distribution and spawns agents in parallel waves:

Before: scaffold → foundations → screen-1 → screen-2 → ... → screen-6 (~35 min)
After:  scaffold → foundations → ┌ components ┐ → ┌ screens ┐ (~18 min)
                                 └ parallel   ┘   └ parallel┘

A new components phase reads all design chunks, builds a component manifest, classifies each (library default, customize, custom, existing), partitions into groups, and spawns agents in parallel. Screen agents then compose from existing components -- no sequential dependency.

Round-robin model assignment (Opus/Sonnet alternating) splits rate-limit pressure across model buckets. Max ~3 agents on the same model simultaneously.

Mixed-Model Critique

Project-critique already spawned two agents (critic + accessibility auditor) but they competed for the same rate bucket. Now the accessibility auditor runs on Sonnet (checklist-based WCAG work) while the critic stays on the user's model. Eliminates rate-limit contention.

Resumable Builds

Each parallel agent writes a status JSON (build/status/*.json) alongside its log. If a session interrupts mid-build, the orchestrator checks for existing complete status files and skips finished agents on resume. STATE.md updates incrementally after each wave, not just at finalize.

Pipeline Gap Fixes

Nine fixes discovered during the parallel orchestration research:

• BUILD-LOG.md race condition -- parallel agents now write individual log files; orchestrator merges after each wave
• Brand sync race -- feedback loop runs synchronously before components phase
• Revision mode compile checkpoint -- QA fixes verified with build command before finalize
• Brand contract enforcement -- critique verdict now checks brand score (X/25) alongside Nielsen (X/50); constraint violations auto-fail
• 4 missing SubagentStop hooks -- accessibility-auditor, brand-auditor, brand-researcher, brand-strategist
• Parallel identity enrichment -- 4 domain skills (logo, color, typography, visuals) invoke simultaneously

Nothing Style Preset

New minimalist design language inspired by Nothing's transparent, dot-matrix aesthetic. Available via /gsp-style --preset nothing.

Performance

Token budget: +171 tokens (+1.1%) over v0.7.0 for ~47% wall-clock time reduction. Red-zone skills unchanged at 4.