skc/tests: batch the invalid-test suite via sharded skc --batch

[mbouaziz]

Summary

Implements #1300 (compiler test harness). The suite runs valid tests in-process via @exptest, but the 826 invalid_tests/ (which must fail to compile) each spawned a fresh skc — re-type-checking the stdlib (~64% of a compile, #1295) 826 times. This batches them through skc --batch (#1298/#1301), so the stdlib is type-checked a handful of times instead of ~800.

Builds on the batch continue-on-error work (#1301) and the frontend-error robustness fix (#1313) that made --batch survive every invalid unit.

How

• compile.sk — skc --batch now writes each failed unit's diagnostics to a per-unit <output>.err sidecar (alongside the existing stderr print), so a driver can attribute failures per-unit without parsing interleaved stderr.
• tests.sk — only the invalid path changes; valid @exptest/fallback handling is untouched.
  • batchCompileInvalid runs in the parent only (sktest workers are re-execs marked by SKTEST_RANK), before test_harness forks the workers. It shards the 826 invalid tests into njobs manifests and runs them as concurrent skc --batch processes — recovering the parallelism the per-test path had, while each shard types the stdlib only once.
  • Outputs go to a deterministic scratch dir under Environ.temp_dir(), recreated each run (no stale artifacts) and derived identically by parent and workers (no coordination needed).
  • Each per-test closure (runInvalidFromBatch) is now a pure filesystem read: an artifact present ⇒ "Expected compile error"; otherwise compare <output>.err to .exp_err (same .trim() as before).

Per-test semantics are unchanged: the .err sidecar holds the same rendered errors a one-shot skc writes to stderr (verified byte-for-byte modulo the trailing newline .trim() already absorbs).

Verified

Built the full compiler test binary with the new harness and ran the whole suite:

• Identical PASS/FAIL vs main: both harnesses report 1723 successes / 1724 on the same machine — every one of the 826 invalid tests and 897 valid tests passes through the batch. (The lone failure, Native CompilerVersionCheck, reproduces identically on main; it's a local multi-stage-build artifact — the test binary bakes the working-tree commit while a locally-built stage1 skc reports its bootstrap commit — and matches in CI.)
• Scratch dir after a run: 826 .err sidecars, 0 binaries, 8 shard manifests — exactly as designed.
• Speed (full skargo test run, same machine): ~58 s vs ~351 s on main — a ~6× reduction, from type-checking the stdlib ~8 times instead of ~800.

Test plan

• full suite: all 826 invalid + 897 valid pass through the batch
• before/after wall-clock on the same machine
• CI

🤖 Generated with Claude Code