docs(plan-reviews): add Execution-settings recommendation preamble

[shanbhagp]

What

Adds a short Execution settings recommendation blockquote at the top of /plan-ceo-review and /plan-eng-review so users pick the right compute tier before invoking the skill.

Execution settings recommendation
Bounded strategic review: assess a plan against business / product /
investor framing. Second-tier compute (e.g., Opus Extra high) is
right. Top-tier compute is warranted only when the plan involves
material strategic ambiguity that benefits from open-ended exploration.

Why

Claude Code now exposes meaningfully different compute tiers (Opus Max, Opus Extra high, Sonnet, etc.). Some skills genuinely benefit from the top tier; many are bounded enough that mid-tier handles them at a fraction of the cost. Without per-skill guidance, users either over-spend on routine reviews or under-spend on review tasks that need headroom.

Wording is task-shape framing ("bounded strategic review", "second-tier compute (e.g., Opus Extra high)") rather than pinned model names, so the guidance survives model-version churn (Opus 4.7 → 4.8 → 5 etc.) without doc maintenance.

The pattern was developed across a set of project-local skills (six narrative-deck skills with similarly-shaped recommendations); the eng + CEO plan reviews were the natural upstream candidates — stable, non-domain-specific, and used widely.

Files

• plan-ceo-review/SKILL.md.tmpl — preamble after H1, before ## Philosophy
• plan-eng-review/SKILL.md.tmpl — preamble after H1, before "Review this plan thoroughly..."
• plan-ceo-review/SKILL.md and plan-eng-review/SKILL.md — regenerated via bun run gen:skill-docs (all three hosts: Claude, Codex, Factory)

Diff is purely additive: 4 files, +26 lines.

Reviewer notes

• Skill-docs freshness gate passes locally — bun run gen:skill-docs && git diff --exit-code is clean.
• Templates and rendered docs are committed together so the skill-docs.yml workflow lands green on first build.
• No version-gate.yml trigger — VERSION / CHANGELOG.md / package.json untouched. Happy to bump if your release cadence wants this in a versioned drop.
• Posture matches your existing template structure (blockquote sits between the H1 and the first body section, not embedded in {{PREAMBLE}} macro space).

Aside (separate issue, not this PR)

While regenerating on Windows I noticed the repo has no .gitattributes, so on Windows clones with core.autocrlf=true (the default), bun run gen:skill-docs produces CRLF-line files where transformFrontmatter()'s strip regex ^${field}:\s*.*\n fails to match — sensitive: true leaks through into Claude SKILL.md outputs. Worked around it locally by setting core.autocrlf=false. Happy to file a separate issue or send a tiny fix PR (a one-line .gitattributes would close it).

[shanbhagp]

One observation that came up while writing this PR — surfacing it here as an offer rather than expanding scope:

The same shape of guidance probably belongs on most of the bounded review skills (slide-review, design-review, plan-design-review, qa-only, review, cso, landing-report). The wording would differ per skill, but the slot is the same: "this is a bounded analysis, second-tier compute is right; top-tier only when X."

I noticed you already have preamble-tier: in the frontmatter for tiered preamble macros. A natural extension would be a compute-tier-recommendation: field that renders uniformly across skills — same content slot, machine-checkable, no per-PR copy-paste drift. But I genuinely don't know your mental model here:

• If you'd prefer that direction, happy to send a follow-up PR proposing the field + renderer changes + rolling it out to the bounded review skills.
• If you'd prefer per-skill human curation (which has its own merits — tighter wording, more judgment per skill), I can expand this PR to cover the rest of the bounded reviewers in the same blockquote style.
• Or merge this as-is and we leave the broader rollout for later — also fine.

No need to decide now; flagging in case it's useful context for review.