Improve MCP agent workflows and evaluation metadata

[NielsRogge]

Summary

• make stock-client MCP 2025-11-25 primary while retaining experimental 2026 discovery; add bounded GET behavior, /.well-known/mcp, generated /docs, initialization guidance, and three prompts
• return compact Markdown fallbacks and compact official-first paper/task/benchmark projections; cap optional resources and expose counts
• add paginated paper evaluations plus task/method search, preserving exact CLI/MCP parity across 21 read-only tools
• merge equivalent leaderboard rows across task scopes and expose rank scopes, evaluation protocol, split, shots, source, openness, timestamps, and metric direction
• prefix actionable failures with not_found, ambiguous, no_markdown, upstream_timeout, or upstream_error; clarify lineage coverage and official/open semantics
• bump pwc-mcp to 0.2.1

Validation

• uv run --with ruff ruff check mcp_server/src/pwc_mcp mcp_server/tests standalone_cli/src/pwc_cli/cli.py standalone_cli/src/pwc_cli/queries.py
• uv run --project mcp_server pytest -q mcp_server/tests (47 passed)
• uv run --project standalone_cli --with pytest pytest -q standalone_cli/tests (82 passed)
• live in-process MCP smoke: 21 tools, 3 prompts, Attention Is All You Need returns 1 official repo plus count 595 in a 64-byte text summary, and MMLU FineRouter returns one model row with Language Modeling and World Knowledge rank scopes

Companion routing PR

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

Supersedes #12, which predated the merged CLI-parity architecture.