`pnpm try` crashes the agent run with `require is not defined` (local dev / ESM)

What's happening

Running the wizard from source with pnpm try (which is tsx bin.ts) crashes once the agent run starts, with require is not defined.

The culprit is getClaudeCodeExecutablePath() in src/lib/agent/agent-interface.ts. It uses require.resolve('@anthropic-ai/claude-agent-sdk') to find the bundled Claude Code CLI. Under tsx we're running native ESM, where there's no global require — so the call throws.

Why prod is fine (and this only bites local dev)

The published build is ESM (format: 'esm' in tsdown.config), but rolldown rewrites require.resolve into a working __require.resolve at build time. So real users were never broken — this is purely a pnpm try / raw-tsx thing where that build-time rewrite doesn't happen.

The fix

Make getClaudeCodeExecutablePath() work under both module systems: use require.resolve when require exists, and fall back to walking up from the entry script to find @anthropic-ai/claude-agent-sdk in a sibling node_modules when it doesn't.

A working version of exactly this fix already exists in #514 (it rode along with the WizardAmp visualizer). The plan is to pull it out of that PR and land it on its own, with a test, so it gets reviewed on its own merits — this issue tracks that.

Acceptance

• pnpm try --install-dir=<path> reaches the agent run without require is not defined
• Production (compiled) path still resolves the CLI correctly
• A test covers the ESM resolution fallback

Start in src/lib/agent/agent-interface.ts at getClaudeCodeExecutablePath(), then inspect the working version in PR #514. Run pnpm try --install-dir= to reproduce the ESM failure and verify the production path still works. Add a test covering the ESM resolution fallback and confirm the agent run starts without require is not defined.