diagnosing-bugs: disclose phases progressively to enforce the feedback-loop gate

## Problem An agent loaded the complete `diagnosing-bugs` skill, then jumped from CI logs directly to a root-cause hypothesis before establishing the required red-capable local loop. The skill already says not to do this, so adding another warning is unlikely to change behavior. Keeping every later phase visible appears to pull the agent toward premature completion of Phase 1. ## Proposed change Use progressive disclosure by sequence: - keep the overview and the current phase's completion criterion in `SKILL.md`; - move later phases into referenced files that are loaded only after the current criterion is met; - make each transition require the command and redacted output proving the previous phase completed; - preserve the existing diagnosis method and completion criteria rather than weakening or expanding them. This follows the repository's own `writing-for-agents` guidance: split a sequence when visible post-completion steps tempt the agent to rush the current one. ## Acceptance criteria - Phase 2 instructions are not loaded until a tight, red-capable command has been run and reported. - Every later phase has an equally explicit entry criterion. - A behavior evaluation demonstrates that the agent does not hypothesize before producing the Phase 1 command and output.