Fix show_neighborhoods tool call reliability

[AndrewSepic]

Summary

While testing this demo, the agent would frequently write full neighborhood recommendations in chat text but call show_neighborhoods with an empty or missing neighborhoods array, so nothing ever rendered on the map. Diagnostic logging showed the model duplicating its answer (full prose, then a redundant tool call) and getting no feedback when that second call came back empty.

• Move the show_neighborhoods instruction to the top of the system prompt, apply it to every follow-up turn, and tell the model to keep chat text brief and put recommendation detail only in the tool call's reason field instead of duplicating it in prose.
• Have execute() return a corrective error (asking the model to retry) instead of a canned success message when neighborhoods is empty or invalid.
• Add logging (raw tool args, per-step tool inputs, and a client-side warning on unexpected tool-part states) so failures are visible instead of silent.
• Fix two pre-existing eslint config gaps surfaced while linting these files: add a node env override for server.js (the first Node/Express file in this repo) so process.env isn't flagged as undefined, and turn off react/prop-types for .ts/.tsx files since this codebase uses TypeScript types instead of PropTypes.

Test plan

• Ran the app locally and confirmed the agent reliably calls show_neighborhoods with non-empty data on first ask and on follow-up refinements
• npx eslint clean on the modified files
• Chris to review/merge into cw/demo-nyc-neighborhood-finder

🤖 Generated with Claude Code