docs: separate execution quality from trigger quality in eval guidance

[christso]

Objective

Explicitly document the distinction between execution-quality evaluation ("does the skill help when loaded?") and trigger-quality evaluation ("does the system load the skill when it should?") in AgentV's docs and roadmap, preventing overloaded evaluation configs and clarifying scope for future work.

Architecture Boundary

docs-examples

Context

Both Anthropic's skill-creator and Tessl treat trigger optimization as a separate evaluation track from task execution quality. Anthropic's skill-creator has dedicated tooling for trigger evaluation: repeated trigger trials, train/test splits, held-out model selection, and description optimization.

AgentV's current eval guidance naturally emphasizes execution quality. Without explicitly calling out trigger quality as a separate concern, users may try to overload execution eval configs with trigger-detection logic, or expect AgentV to handle skill discovery optimization that belongs in a different product surface.

Design Latitude

• Choose where this distinction is documented (roadmap doc, architecture doc, existing eval guide, or new conceptual guide)
• May be as simple as a "Concepts" or "Evaluation Types" section in the docs
• Choose how to frame trigger-quality as a future direction without over-promising
• May reference the skill-creator's trigger evaluation approach as industry context

Acceptance Signals

• AgentV docs explicitly name "execution quality" and "trigger quality" as distinct evaluation concerns
• The docs explain why they are different problems (noisy vs deterministic, different optimization surfaces)
• Current AgentV eval tooling is positioned as execution-quality evaluation
• Trigger-quality evaluation is framed as a future direction, not a current gap
• The docs include a clear statement that execution eval configs should not be used for trigger evaluation
• The agentv-eval-builder skill reference card is updated to reflect this distinction (per CLAUDE.md Documentation Updates guidelines)

Non-Goals

• Building trigger-evaluation CLI commands or runtime features
• Adding trigger trial tooling, train/test splits, or description optimizers
• Creating a skill marketplace or discovery system
• Changing the current eval schema or config format

Research Basis

• Skill Creator Findings — Key Pattern #5: Trigger optimization is a separate eval track
• Tessl Findings — Key Pattern #5: Separate skill execution quality from skill discovery/trigger quality
• Skill Lifecycle Alignment — Adopt Now #3