feat(cu-benchmark): reusable CU benchmark report action

[dev-jodee]

What

A composite action that turns a freshly generated compute-unit (CU) report from a Solana program's tests into a PR comment with per-instruction Avg CU deltas vs main, and can optionally keep a committed baseline in sync.

Useful for any program repo that wants visible CU regressions on every PR without extra tooling.

Behavior

1. Auto-locate cu_report.md in the checkout (fails loud on 0 or >1; report-path overrides). Baseline = its sibling cu_baseline.md.
2. Read the baseline from main; diff the fresh report against it.
3. Upsert a PR comment with a Δ Avg vs main column (🔺/🔻/–/🆕/🗑), deduped by a hidden marker.
4. Opt-in: on same-repo PRs, fetch+switch to the head branch, commit the refreshed baseline, push. Loop-safe (pushes via GITHUB_TOKEN don't re-trigger workflows). Fork PRs skip.

base=main; head ref / PR number / token are derived from the github context in-step — not inputs.

Inputs

input	required	default	notes
report-path	no	"" (auto-discover)	override only when a repo emits several cu_report.md
commit-baseline	no	false	opt-in auto-refresh; needs contents: write + a same-repo check

Permission profiles

Comment-only (default) — needs only pull-requests: write. Deltas work against a committed baseline; with no baseline, every instruction shows 🆕.

Auto-refresh baseline (opt-in) — add contents: write and set commit-baseline to the same-repo check.

Example caller workflow

name: Benchmark
on: pull_request
permissions:
  contents: write        # only if commit-baseline enabled
  pull-requests: write
jobs:
  compute-units:
    runs-on: ubuntu-latest
    timeout-minutes: 30
    steps:
      - uses: actions/checkout@v6
        with: { fetch-depth: 0 }
      - uses: ./.github/actions/setup        # your toolchain setup
      - run: just test-and-benchmark          # produces cu_report.md
      - uses: solana-developers/github-actions/cu-benchmark@vX
        with:
          commit-baseline: ${{ github.event.pull_request.head.repo.full_name == github.repository }}

Verification

Diff/render tested against tabled-style tables (increase, decrease, unchanged, new, removed, empty baseline, mismatched columns). Auto-discovery tested for 0 / 1 / multiple matches + override. Portable bash (no bash-4 mapfile).

Notes

• base is hardcoded to main: stacked / release-branch PRs diff vs main, not the parent branch.
• Parses the report's markdown table by column header; a future cu_report.json companion could make diffing format-agnostic.