ci: drop the AppVeyor Node matrix - one Windows job, shallow clone

[yagop]

Why

AppVeyor was taking 4-10 minutes end-to-end on recent PRs (e.g. #1332) for ~65 seconds of actual work. Timing data from the AppVeyor API for the last 15 builds:

• Queue wait: 1-6.5 min before the first job starts (free-tier shared queue - not fixable via config).
• Run: ~190-240s, of which each matrix job is only ~60-65s. The free plan runs 1 concurrent job, so the Node 22 + 24 matrix ran sequentially, with a second ~65s VM provision between them.

Per-job breakdown (build 357): clone 8s, Install-Product node 10s, npm install 22s, typecheck 5s, unit tests 18s.

What

• Drop the Node matrix, keep a single Node 24 job. This job exists for Windows coverage only; the Node version matrix (22/24/26) is already covered by GitHub Actions on Linux, where each job finishes in ~20s. Cuts the AppVeyor run from ~3.5 min to ~1 min and halves queue exposure (one job queues instead of two).
• clone_depth: 1 - shaves most of the 8s full clone.

What was considered and not done

• Caching node_modules: would save ~15s of the 22s install, but cache save/restore on AppVeyor costs a similar amount for a tree this size, and there is no npm lockfile (repo ships bun.lock) to key it reliably. Not worth it.
• The queue wait itself can't be configured away. If 1-6.5 min of queue is still too slow, the real fix is moving Windows coverage to GitHub Actions (runs-on: windows-latest starts in seconds) and retiring AppVeyor - happy to do that in a follow-up if preferred.

🤖 Generated with Claude Code