feat(skills): add /ponytail-gain measured-impact scoreboard

[ousamabenyounes]

What

Adds /ponytail-gain — a one-shot scoreboard that renders ponytail's measured benchmark impact (less code, less cost, more speed) as plain ASCII bars, then points at /ponytail-debt and /ponytail-audit for this repo's real numbers.

  ponytail gain                     benchmark median · 5 tasks · 3 models

  Lines of code   no-skill  ████████████████████  100%
                  ponytail  ██▌·················    6–20%   ▼ 80–94%
  Cost            no-skill  ████████████████████  100%
                  ponytail  █████▌··············   23–53%  ▼ 47–77%
  Speed           ponytail  ▸ 3–6× faster

  This repo:  /ponytail-debt  (shortcuts you deferred)
              /ponytail-audit (what's still cuttable)

It complements the existing skills instead of duplicating them: debt harvests the ponytail: ledger, audit finds what's cuttable, gain shows the measured why it matters. The first thing a teammate trying ponytail asks is "what does this actually buy me?" — this answers it in one command, and routes them to the per-repo skills for the specifics.

Our position on "tokens saved" (and why this is light, on purpose)

We really wanted gain to report real token/LOC savings on your repo — the rtk-gain style number. We talked ourselves out of it, twice, and want to be transparent about why:

1. No honest baseline in a live repo. ponytail changes what the model generates; the bloated 400-line version it didn't write was never captured. Claiming "you saved 380 lines here" means subtracting from a file that never existed — an invented number. That's exactly the kind of unmeasured figure this project refuses to ship.
2. Live per-read/edit tracking is too costly and unportable. Measuring the maintenance dividend (a smaller file is cheaper every time it's read) would need a hook firing on every file Read/Edit, a growing state file, and a tokenizer — a PostToolUse-style sys-call on each tool use. That breaks the zero-config, thin-adapter promise (Cursor/Windsurf/Cline have no such hook) and contradicts "Does it need a config file? No."

So we cut it down to light-only: the bars carry the published benchmark medians (measured, labelled as benchmark not as your repo), and the only per-repo figures come from debt's real count. No phantom numbers, no runtime weight, portable everywhere. If you'd rather measure the maintenance/lifetime cost properly, the honest place is a new arm in benchmarks/ (both no-skill and ponytail files already exist there) — happy to follow up with that if you want it.

Adapter coverage

Ships every surface the other commands ship, so no adapter drifts:

• skills/ponytail-gain/SKILL.md, commands/ponytail-gain.toml
• .opencode/command/ponytail-gain.md
• .openclaw/skills/ponytail-gain/SKILL.md (generated via scripts/build-openclaw-skills.js)
• Pi command registration
• Help card, /ponytail-help enumeration, docs/agent-portability.md, and the README commands table + OpenClaw line updated in the same PR.

Test verification (RED → GREEN)

CI is green on this PR (the test workflow passed). The Pi command-registration + delegation tests are the anchor — they go red the moment the prod registration is reverted, green with it.

RED — prod registration reverted to upstream, test kept:

$ git checkout origin/main -- pi-extension/index.js && node --test pi-extension/test/*.test.js
not ok 1 - extension registers Ponytail commands
    + actual - expected
    -   'ponytail-gain',
  name: 'AssertionError'
not ok 4 - skill alias commands delegate to Pi skill commands
# fail 2

GREEN — with the change:

$ node --test pi-extension/test/*.test.js
# tests 11
# pass 11
# fail 0

$ node scripts/check-rule-copies.js
Rule copies match AGENTS.md; 5 rule invariants present in SKILL.md and AGENTS.md.

Baseline preserved. Note for anyone running the suite locally: tests/correctness includes one check (csv: correct pandas one-liner) that imports pandas, so node --test tests/*.test.js shows one failure on a machine without it. That is environment-only and not caused by this change — CI installs pandas (pip install pandas in .github/workflows/test.yml), which is why the workflow is green. It was red the same way on main before this branch.