perf(pairing): trim always-on frontmatter and dedupe body prose

[liwenjie200543]

Summary

• Applies the Optimize skill token cost across all families (setup-family recipe) #1342 setup-family optimization recipe to the pairing skill family (Optimize the pairing skill family #1352). Both pairing skills were over the 200-token always-on budget — description + when_to_use are paid in every session, for every skill, invoked or not.
• pairing-multi-agent-review: always-on ~230 → ~172 tokens (chars÷4 estimate); measured body tokens 3,686 → 3,467.
• pairing-self-review: always-on ~209 → ~170; measured body 3,437 → 3,377.
• The body pass dedupes prose that repeated the frontmatter description, the per-pass scopes, and golden-rule bodies. Headings, golden-rule headlines, code blocks, and the pre-flight block are byte-identical; all quoted trigger phrases in when_to_use are preserved.

Test plan

• uv run --project tools/skill-token-count skill-token-count — stamps current, exit 0 (restamped via --write in the same PR).
• uv run --project tools/skill-and-tool-validator --group dev skill-and-tool-validate — exit 0. The only warnings (override-contract, no-telemetry-import) are pre-existing on files this PR does not touch.
• python tools/dev/skill-surface-hash.py — reports "surface_hash in sync across 78 skills", i.e. all structural anchors are unchanged.
• Eval-coupled strings verified present (e.g. the Pass A injection-guard finding text).
• The behavior eval suites for both skills (15 + 20 cases) require claude -p and were not run in my environment — noted here per the Optimize skill token cost across all families (setup-family recipe) #1342 recipe's "run evals before and after" step. Happy to run them locally on request, or maintainer run appreciated before merge.

Gen-AI disclosure

Prepared with AI coding assistance (ZCode); changes reviewed, tested, and committed by me.

[Kaap10]

Thanks @liwenjie200543!

Trimming frontmatter on skills is always tricky because it's easy to accidentally drop subtle trigger phrases or prompt context, but this is done really cleanly.

Checked the changes across both multi-agent-review and self-review:

• Both frontmatters are comfortably under the 200-token budget now without losing any of the when_to_use trigger keywords.
• The prose deduping in the body keeps the 3-axis isolation model, Step 1–4 pipeline, and Golden Rules completely intact.
• Measured token counts match.

[onlyarnav]

lgtm