Simplify the tool surface and add meal ratings

[colinds]

Adds rate_meal so members can submit or edit a 1–5 meal rating, reasons, comments, and per-rating preferences through the existing preview and confirmation gate. Ratings target an exact owned piece, preserve omitted feedback and existing attachments, and include historical lookup arguments when an outcome is uncertain. Delivery lists and status responses now show owned rating feedback. Historical lookups accept both from and to, score-change previews show the old and new score, and reason normalization preserves unknown server codes while removing known incompatible codes and duplicates.

Removes skip_delivery and explain_pick, bringing the server to 11 tools. Skipping uses list_deliveries plus remove_meal; recommendations replace pick comparisons without preserving exact ranks outside the returned suggestions. Deletes unused configuration, result-conversion, transport, selection, and timestamp code, trims GraphQL selections, and shares ownership filtering so a missing user ID never claims another member's meal. Keeps the 20-minute keepalive and existing mutation recovery safeguards.

Validation:

• bun run check: lint, formatting, type checking, and 226 tests passing in each timezone run.
• bun run smoke: packaged installs pass with both Bun and Node, each exposing the expected 11 tools and rating schema.
• Rating HTTP requests, null/URL attachments, feedback preservation, reason normalization, duplicate targets, ownership, confirmation reuse, exact rejection messages, and bounded historical reconciliation are covered with mocked responses. The wire contract follows the locally captured Forkable dashboard. Live connected-MCP checks verified the 11-tool surface, profile/delivery/menu/search/recommendation reads, modifier and date guards, and meal-operation previews. A user-authorized rating submission succeeded and was independently read back through both delivery read tools; upcoming meals and confirmation states remained unchanged.

Buffet ratings, photo editing, and account-setting changes are not included.

The live rating check showed that Forkable can enable rating follow-ups when the prior preference is unreported and the argument is omitted. The preview and documentation now label that as Forkable’s default rather than claiming the preference stays unchanged. Feedback-clearing and existing-photo preservation remain covered by mocked tests only.