ci: speed up Windows integration tests

[lwshang]

Problem

The Windows test jobs spend ~10 min compiling in the "Run tests" step while Ubuntu takes 1–2 min — and it's been this way regardless of PR. The Rust cache is working, but it only stores third-party dependency artifacts. By design Swatinem/rust-cache strips workspace crates from target/ before saving (cache-workspace-crates: false), and the per-file integration test binaries (one per tests/*.rs) plus linking are never cached. So every test job recompiles the whole workspace and links its test binary from scratch. That residual work is ~1–2 min on Ubuntu but ~10 min on Windows because of slow MSVC link.exe linking, full debuginfo, and Defender real-time file scanning.

Changes (quick, CI-scoped — no structural change)

All in .github/workflows/test.yml:

• CARGO_PROFILE_DEV_DEBUG=line-tables-only — drop full debuginfo so object files are smaller and link faster; keeps file:line in backtraces. The test profile inherits this.
• lld-link on windows-msvc — link with LLVM's lld-link (preinstalled at C:\Program Files\LLVM on windows-2025, no install step) instead of the slow link.exe. Only consumed when building the windows-msvc target → no-op on Linux/macOS. Scoped to this workflow, so the release/dist thin-LTO build keeps the default linker (lld + LTO has known segfault caveats).
• Windows Defender exclusion for the workspace + .cargo/.rustup, in both the unit-tests and test jobs. Non-fatal so it can't break CI.

Notes for reviewers

• First run recompiles fully on every OS — changing the linker and debug level invalidates the cache fingerprint once. Steady-state is what improves.
• The lld-link swap is validated by this PR's Windows CI run (can't link MSVC locally from macOS). If lld-link.exe isn't found at that path on some image variant, those jobs fail at link with an obvious "linker not found" error; the fix/revert is a one-line env change.
• Deferred follow-up: the larger structural win — prime all integration test binaries once in the unit-tests job + cache-workspace-crates: true so the parallel test jobs reuse pre-built binaries instead of each recompiling the workspace. Larger payoff but grows caches and needs eviction-limit care.

Verification

Confirmed against this PR's Test run on Windows:

• Warm-cache compile+link dropped from ~10 min to ~3.2 min. Example: sync_tests spent 194s compiling/linking + 182s actually running tests (was ~10 min compile before). Jobs that hit the deps cache finished in ~380s end-to-end.
• lld-link works — all Windows jobs linked and passed; no "linker not found" errors, so LLVM's lld-link is present and used on windows-2025.
• The few slow jobs were cache misses, not a regression. They cold-rebuilt dependencies (~7 min extra) because the repo is currently over GitHub's 10 GB cache cap and 30 jobs contend for the single cache unit-tests had just created — a transient warm-up effect, unrelated to the linker/debuginfo change. This clears once the new cache key is the norm across main and open PRs, which is why the rollout plan is: merge this, then rebase/merge it into existing PRs so all future CI populates and reuses the new caches.

🤖 Generated with Claude Code