Add stage-awareness rules to Refiner agent

[digitalgrove-bit]

Summary

Adds a mandatory stage-awareness check to the Refiner agent so it stops over-engineering MVPs.

Motivation

In its current form, the Refiner applies SoC and tactical-DDD patterns without regard to project stage, team size, or existing layout. In one real case, it produced a design with:

• 3× file count growth (~12 → ~40 files)
• Two new bounded contexts
• Typed UUID classes
• Aggregate roots over single tables
• CQRS-style read ports

…all for a solo-developer, pre-customer project. The Critique agent rejected nearly all of it (4 CRITICAL + 10 HIGH + 10 MEDIUM + 8 LOW findings). The over-engineering wasn't a Critique failure — the Refiner had no instruction to ask "is this stage-appropriate?" before applying the patterns.

What this changes

Adds a ## Stage-Awareness (MANDATORY before proposing refinements) section to architect-refine-critique/agents/refiner.md with four steps:

1. Read project context — CLAUDE.md and docs/prd.md to extract team size and stage; scan existing codebase layout.
2. Write a Stage Preamble at the top of refinements.md declaring team size, stage, existing layout, and any conflicts.
3. Apply three constraint rules:
   • R-1: No new layout conventions without explicit request (don't retrofit features/platform/shell onto a flat app/)
   • R-2: 50% file-count growth budget — each excess file needs a concrete present-tense justification, not "this scales better"
   • R-3: Match patterns to stage (pre-customer/MVP → flat modules, plain dataclasses, single-file aggregates; scaling → aggregates and value objects; mature → full DDD toolkit)
4. Self-check before writing refined.md.

Default bias: when in doubt about stage, do less. The Critique agent flags under-engineering as readily as over-engineering, and adding structure later is cheaper than ripping it out.

Why this fits the chain

The Architect → Refine → Critique chain works because each agent has a clear job. Adding stage-awareness to the Refiner doesn't change its responsibility — it constrains the search space the Refiner explores. The Critique agent still has the final word; it just has less to push back on when the Refiner stays in scope.

Test plan

• Run /arc on a solo-developer pre-customer project; confirm refinements.md opens with a Stage Preamble and respects R-1/R-2/R-3.
• Run /arc on a project with CLAUDE.md declaring team size; confirm Refiner extracts and uses that signal.
• Run /arc on a project with an existing features/platform/shell layout; confirm Refiner doesn't propose alternative layouts.

Notes

This change has been running locally in production for a week without regression. Happy to iterate on wording or scope if you'd prefer a lighter touch.