cltest: share one go-build cache across captured runs

[cpunion]

CI-time investigation follow-up: the stage-5 branches' mac coverage job was running the cl package at 37-45 minutes, right at the 45m go-test alarm.

What this PR does

RunAndCapture created a fresh temp GOCACHE per test and deleted it afterwards, so each of the ~170 golden-dir run tests re-typechecked the whole dependency graph from scratch. This shares one per-process temp cache instead:

• isolation from the developer cache is unchanged (still a fresh temp dir per test process, removed by the OS temp cleaner);
• sharing is within one test binary only (sync.Once package state) and tests inside a binary run serially, so no new concurrent access surface is introduced across the parallel test processes;
• the go build cache is content+flag keyed, so sharing across build configs is safe.

Measurements (single-variable A/B: this one-file diff vs its parent, no -p flags)

• Local (darwin/arm64, interleaved rounds, heavy run-suites): 348s/308s -> 251s/240s, i.e. 328s -> 245s avg (-25%).
• CI coverage jobs vs the parent commit: mac 1508s -> 1061s (-30%), ubuntu 1376s -> 779s (-43%).
• Full ./cl and ./ssa suites green with the change; this PR's CI is green twice (pre- and post-scope-trim).

What was considered and rejected first

• Removing the ssa TestFrom* suites as duplicates of cl's TestRunAndTestFrom*: rejected — they verify the same golden specs, but per-package coverage attribution means they carry the ssa package's coverage: 87.5% with them, 51.3% without (measured).
• t.Parallel() on the golden subtests: rejected for now — cltest's capture path swaps os.Stdout/Stderr, os.Chdirs, and mutates GOCACHE globally, and LLVM installs in-process signal handlers; under a cold cache this reproducibly fails (including a hard fatal error: semasleep on Darwin signal stack). Parallelizing needs a capture refactor (explicit cmd.Dir/Env/Stdout), tracked separately.

Scope note

This PR is the cache sharing only. On the much heavier stage-5 branches, warm caches let the four heavy test packages' compile phases overlap fully, and the concurrent clang/lld burst killed 16GB ubuntu runners mid-run; those branches carry a go test -p 2 mitigation locally. That cap is deliberately NOT part of this PR — it lowers the collision probability rather than provably removing the peak; the structural fix (splitting heavy packages into separate CI jobs) is the tracked follow-up if the deaths reappear.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!