ci: share a bench setup action and stop evicting the browser cache

[harshtandiya]

Playwright job usually ran ~5 min but spiked to ~15. The variance was one step: Install Playwright, 16s on a good run vs 624s on a bad one (runs 30474649946 / 30473372443). Tests themselves are stable at ~110s.

Cause: gh cache list shows the repo at GitHub's 10 GB cache cap with >8 GB of it yarn (Linux-yarn-* at 1.17 GB, 1.18 GB, 0.90 GB ×2, 0.77 GB). LRU eviction kills the 249 MB Playwright cache, so chromium re-downloads. The yarn caches grew forever because restore-keys restored the previous cache, bench build added to it, and the union got saved under a new key.

Fix: both test workflows now use setup-node's built-in yarn cache, no restore-keys, sharing one entry with typecheck/unit-tests. Playwright cache keys on yarn.lock instead of hashFiles('**/package.json') — dashboard/package.json changed 5× in the last 50 commits and was busting it for unrelated reasons.

Also extracts the provisioning both workflows duplicated into .github/actions/setup-bench, and overlaps what shares no state: apt with uv tool install frappe-bench, and bench build with bench new-site. Plus mariadb 10.6 → 11.8 with tmpfs and durability off, and bench build dropped from the server job.

Non-obvious bits a reviewer can't get from the diff:

• bench build runs root package.json's "build": "cd dashboard && yarn build" via frappe's esbuild --run-build-command. It's a full vite build — that's why overlapping it with new-site is the win, not a micro-opt.
• No NODE_ENV: production, though frappe sets it. vite, tailwindcss, @vitejs/plugin-vue and @playwright/test are all devDependencies here; it would break the build and the tests.
• socketio stays in the Procfile. Frappe only strips it in server tests. dashboard/src/main.ts:44 calls initSocket() on boot.
• Root install needs --ignore-scripts — the root postinstall is cd dashboard && yarn install.
• bench setup requirements --dev installs only python dev deps (bench/commands/setup.py, the else branch), so running it alongside --node is not a race.
• bench get-app's NAME arg is nargs=-1 and documented "Dummy argument for backward compatibility" — a bare URL is enough.

Not done: sharding Playwright (110s of tests vs ~100s provisioning per shard), workers > 1 (projects share server state, playwright.config.ts:20), caching all of ~/frappe-bench (1–2 GB into the budget that caused this).

Deferred: the oversized Linux-yarn-* cache entries still need gh cache delete. Pointless before this merges — develop's current workflows would recreate them on the next run.

Checks: actionlint clean, YAML parses, shellcheck clean on every run block, composite inputs cross-checked against both callers. Not run: the workflows themselves — they only execute on GitHub, so CI on this PR is the first real test. Expect the first run to be slow (cold caches); the second is the measurement.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.