docs(skill): add "keep related scenarios in one tree" guidance

[stepan-romankov]

What

Adds a new sub-section to the samurai skill's "File layout" guidance: Keep related scenarios in one tree.

When one scenario mutates state another reads back (grant → read-shows-present vs. revoke → read-shows-absent), consolidate into a single RunWith tree instead of fragmenting one-file-per-action. Independent features still get their own files.

Why this wording

Chosen via an A/B eval — 5 skill configs (baseline + 4 candidate wordings) × 3 prompts (write-from-scratch, review-and-defragment, independent-features-boundary):

• The section works: the discriminating prompt was review-defragment. Baseline (no section) read three single-leaf fixture files and removed samurai entirely; all four wordings instead recognized the shared mutating prefix and merged into one tree.
• Behavior was tied across all four wordings — all consolidated correctly, none over-merged the independent features (boundary clause holds).
• The decision-first phrasing won the tie-breaker: leanest (100 words vs. 130 for the prose candidate), opening with the exact question the reader must answer. Token cost across configs was within ~1.5% noise.

🤖 Generated with Claude Code