fix: ship a node-runtime CLI to actually fix stdout truncation (#20)

[stephendolan]

Summary

v2.5.1 removed process.exit() calls but issue #20 reproduced anyway. Root cause is the #!/usr/bin/env bun shebang — bun drops queued stdout writes on exit even without an explicit process.exit(). The same bundled dist/cli.js:

• under node: 10385 bytes piped ✅
• under bun (current shebang): 512 bytes piped ❌

This PR:

• Switches the shipped shebang to #!/usr/bin/env node (tsup emits plain ESM; no bun-runtime APIs in the bundle).
• Drops engines.bun, adds engines.node: ">=20". Bun remains a dev dependency for TS execution, build, and tests.
• Adds program.exitOverride() so commander's --help, --version, and parse-error paths don't call process.exit and race the pipe drain. The CommanderError.exitCode is forwarded via process.exitCode.
• Adds src/lib/__tests__/pipe-truncation.test.ts, an end-to-end regression suite that spawns the built dist/cli.js under its real shebang and counts bytes received through a real cli | cat pipe. The v2.5.1 test passed because the error payload fit in one 512-byte pipe buffer — this one uses --help (~649 bytes) so the regression would now be caught.
• CI: build before test so the integration suite has dist/cli.js to exercise.
• CLAUDE.md: document the runtime split and the exitOverride() requirement.

Test plan

• bun run typecheck clean
• bun run lint clean
• bun run test (vitest) — 24/24 pass
• bun test (bun native, mirrors CI) — 24/24 pass
• Manual: dist/cli.js inbox list 2>/dev/null | cat | wc -c → 10385 (was 512)
• Manual: dist/cli.js --help | cat | wc -c → 649 (was 512)
• Manual: error exit code 1 still propagates after the change

🤖 Generated with Claude Code