Apply recommendation-gradient universally in interactive sessions

[williamthorsen]

Problem

The recommendation-gradient format is referenced only from the collaboration and design-and-plan skills. Other skills that present options (e.g., refine-plan, plan-orchestrable-steps, brainstorming, assess-ticket) and ad-hoc option-asking outside any skill don't trigger it. Agents apply the format inconsistently, missing its productivity benefit.

Two related friction points surface in practice:

1. The format specifies a 3-space indent for ➕/➖ continuation lines, but agents commonly flatten leading whitespace, so pros and cons appear unindented and visually merge with the option title.
2. Multi-question responses give the user no way to reference questions unambiguously ("for option 2 of which question?").

Solution

1. Promote the directive to top-level guidance. Add a short rule to the shared AGENTS.md (packages/agents/content/guidance/shared/AGENTS.md) anchoring all interactive sessions: when presenting 2+ numbered options to the user, follow _data/recommendation-gradient.md.
2. Single-source the rule. Remove the gradient references from collaboration/SKILL.md and design-and-plan/SKILL.md.
3. Switch indent to NBSP (U+00A0). Update _data/recommendation-gradient.md's format spec and examples to use 3 NBSP characters instead of 3 ASCII spaces. Visible characters survive whitespace normalization in model output.
4. Add question identifiers. When a response contains 2+ option-style questions, prefix each with Q1, Q2, etc., so the user can reference them unambiguously.

Acceptance criteria

• Shared AGENTS.md contains the recommendation-gradient directive
• References removed from collaboration/SKILL.md and design-and-plan/SKILL.md
• _data/recommendation-gradient.md format spec and examples use NBSP for indentation
• _data/recommendation-gradient.md documents the Q1/Q2 question-ID convention