test(harness): run one conformance suite on every sandbox driver - #77

Summary

• New shared conformanceSpec (test/fixtures/conformance.spec.ts), run once per driver: host, memory, sqldb in test/sandbox.conformance.test.ts, and Vercel, Daytona from their E2E files on the sandbox each one already provisions. memory and sqldb are just-bash over a VFS, so they also cover just-bash.
• Cases:
  • fs: read/write round-trip, overwrite, parent creation, read errors, buffers and byte ranges, scanLines corpus, stat/lstat incl. symlinks, readdir, exists, mkdir, rm
  • paths: relative → cwd, .. normalised, .. clamped at /
  • fs ↔ shell: each sees the other's writes
  • shell: exit codes and streams, per-command env, cwd, compound commands, pipelines and find, large output through ToolShell, typed timeout, pre-interrupted commands never start, mid-flight interrupt
  • host only: a child that ignores SIGTERM is killed
• Each driver writes test/__artifacts__/sandbox.conformance.<driver>.json. Driver differences are declared options, never skipped: combinesOutput (Daytona), cancels (Daytona: false), inheritsHostEnv (host).
• Replaces the old filesystem spec, the Flue parity matrix and the cancellation spec. remote.spec.ts keeps reattach/persist, GitHub push, and a check for the provider's baked env. withService moves to fixtures/service.ts.
• The remote lifecycle E2E now calls wake and checks that create and wake each record codework.sandbox.provision.duration once.

Daytona fixes found by the suite

• A non-recursive mkdir created missing parents, because the toolbox's createFolder is mkdir -p. It now runs mkdir in the shell.
• rm of a missing path was not reported as not found. It now returns ENOENT.

Test plan

• vp check, vp run typecheck
• vp test test/sandbox. — 142 passed
• Live conformance, reattach and push on Vercel and Daytona: 42 passed
• sandbox.remote.driver.e2e.test.ts (provision metric): passed
• Live bashPluginSpec and runner cycle on Vercel and Daytona: 13 passed

🤖 Generated with Claude Code