[neilmsft]

Describe your changes

Add a manually invoked experimental olive-pr-review Skill under skills/dev.
It produces private, evidence-backed advisory reviews limited to provisional H01
(capture CLI intent/config propagation) and H03 (explicit CompositeModel selection).
Include scoped exceptions/test-evidence limits, a structured output template, trusted
loading instructions, and two practical architecture walkthroughs.

No Olive runtime changes, automatic workflow, GitHub posting, approval, or merge behavior.
No dependency on the unmerged architecture draft. Maintainer confirmation remains pending.
Keep deferred config duplicate rejection distinct from immediate handler rejection.
Research, synthetic development inputs/labels, private outputs, and local tooling are
intentionally outside this contribution.

Validation:

• git --no-pager diff --check — passed.
• lintrunner skills\README.md skills\dev\olive-pr-review\SKILL.md skills\dev\olive-pr-review\README.md skills\dev\olive-pr-review\references\rules.md skills\dev\olive-pr-review\references\review-template.md — passed all applicable configured checks. Windows required existing Git grep/sed in process-local PATH; no checks disabled. Python checks are not applicable to these Markdown files.
• copilot skill add / copilot skill list / copilot skill remove on a private trusted directory — discovery verified and registration cleaned up.
• copilot -C $pilot --available-tools view --disable-builtin-mcps -p $prompt — passed in three fresh explicit-loading demonstrations. $pilot and $prompt identify private development inputs; all three runs read the Skill/references before candidate evidence. Detected synthetic H01/H03 issues and avoided a finding on a valid refactor/validation-timing case; expected labels were separate.
• Independent private checks verified source/diff hashes against the pinned base and public-file boundaries. The examples are synthetic proposed heads, not real PR replay or benchmark accuracy evidence.
• Olive/pytest/model/runtime tests and live PR/real commit-pair integration were not run. No historical or candidate code was executed. Read-only demo reviewers reported independent hashing/freshness limits.

Contributor-facing release note: experimental private H01/H03 review Skill with explicit
loading and architecture examples; advisory only, no approval or publication.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

Checklist is intentionally left for maintainer/manual review. This Markdown-only prototype
adds no runtime behavior or unit tests; executed checks and skipped runtime validation are
listed above. Lint ran without auto-fix and found no changes to apply.

(Optional) Issue link