fix(cli): stop a run's created sandbox before the harness closes - #75

Summary

• codework run exits 0 and stops the sandbox it created. The run's scope wrapped the provided harness layer, so the finalizer that stops a created sandbox ran after the database had closed. It died with statement has been finalized (a defect, so Effect.ignore did not swallow it) and the run exited 1 after printing the response. Daytona and Vercel sandboxes were left running, because the stop never reached the provider. Effect.scoped now sits inside the provide.
• codework run logs to stderr, like acp, so stdout carries only the response.
• Daytona reports missing paths as ENOENT. The toolbox rejected with DaytonaFileNotFoundError, which isNotFoundError does not recognise, so every missing AGENTS.md/CLAUDE.md logged a warning with a full axios dump. This also gives the read tool its usual not-found error on Daytona.
• Tests load credentials from the root .env.local (aikit, codework, harness and the Vercel smoke script).
• Runner-cycle E2E no longer races the interrupt. A 100-word reply could finish streaming before the interrupt landed (expected 'ended' to be 'failed'). It now asks for a reply long enough that it cannot.

Testing

• New packages/codework/test/cli.sandbox.live.test.ts spawns the real CLI on memory, Daytona and Vercel sandboxes (OpenRouter deepseek/deepseek-v4.1-flash). It asserts exit 0, no ERROR/WARN on either stream, and snapshots the final sandbox row (offline, provider stopped) to __artifacts__/cli.sandbox.live.<driver>.json. It fails on the old run.ts (exit 1) and on the old Daytona provider (WARN lines).
• Harness remote E2E (CODEWORK_SANDBOX_E2E_REQUIRED=1: vercel, daytona, remote.driver, tools.registry.remote, session.relink, plugin.instruction): all pass. The runner-cycle case passed 3/3 after the fix.
• codework cli.test.ts passes.
• vp check and vp run typecheck pass.
• Not run: sea.test.ts, which is not on this branch yet.

🤖 Generated with Claude Code