Improve MCP agent experience and catalog coverage

[NielsRogge]

Summary

• return typed catalog failures with ambiguity candidates and fix exact-title, related-paper, empty-result, timeout, and Markdown error handling
• compact text fallbacks and resource-heavy tool outputs while preserving structured content; cap paper resources and task benchmarks
• add trending papers, paper evaluations, task/method enumeration, official-first resources, Hugging Face artifacts, and paginated leaderboard/evaluation results
• deduplicate equivalent model rows while preserving task-scoped ranks; expose split, shots, protocol, source, metric direction, openness, timestamps, and lineage coverage caveats
• advertise stock-client MCP 2025-11-25, add generated /docs and /.well-known/mcp, return 405 for bare GET /mcp, and add initialization guidance and example prompts
• expand schemas/descriptions with IDs, slugs, dates, filter semantics, and absolute Papers With Code URLs; bump pwc-mcp to 0.2.1

Validation

• uv run --project mcp_server pytest -q mcp_server/tests (39 passed)
• uv run --project standalone_cli --with pytest pytest -q standalone_cli/tests (76 passed)
• uv run --with ruff ruff check mcp_server/src/pwc_mcp mcp_server/tests
• uv run --project standalone_cli python standalone_cli/scripts/check_contract.py
• live catalog smoke tests for title lookup, related/trending papers, paper evaluations, Image Classification task response bounds, Mamba, ImageNet, text-to-video ambiguity, and MMLU rank scopes

Companion routing PR

The human setup and canonical discovery routes are in huggingface/paperswithcode#266.

[NielsRogge]

Superseded by #17, rebased cleanly on the merged CLI-parity and configurable-rate-limit architecture. The replacement retains these agent-experience fixes and adds the follow-up evaluation/ranking changes.