fix: stabilize E2E tests against Bun pipe buffering and LLM non-determinism

[FL4TLiN3]

Summary

• Fix stdout capture in E2E test runner by redirecting to temp file instead of pipe, working around Bun's pipe buffering that truncated large outputs (6.3MB → 1.5MB)
• Rewrite NDJSON event parser with boundary grouping to handle base64 data containing literal newlines inside JSON strings
• Relax event sequence assertions in PDF/image tests to tolerate LLM batching tools in fewer turns
• Fix dynamic skills test: use local bun ./apps/base/dist/bin/server.js instead of npx @perstack/base (which failed with "could not determine executable to run" since addSkill bypasses the bundled base optimization)
• Add runCliUntilToolCalled retry helper for tests sensitive to LLM non-determinism
• Strengthen expert instructions in multi-modal and skills TOML configs
• Add explicit preservation requirement for existing experts in create-expert definition-writer

Test plan

• run.test.ts — 4/4 pass, 3 consecutive green runs
• skills.test.ts — 6/6 pass, 3 consecutive green runs
• create-expert.test.ts — 2/2 pass, 3 consecutive green runs

🤖 Generated with Claude Code