feat(papers): live on-demand 'search recent papers' via opencite

[neuromechanist]

Summary

Implements the live paper search idea (#308): a tool that fetches the most recent literature on demand via opencite, beyond what the batch sync has indexed. Builds directly on the opencite client added in #309.

Closes #308.

What's added

• search_<community>_papers_live — a new agent tool. It queries opencite across (openalex, s2, pubmed) sorted newest-first, returns the top results, and best-effort caches them into the community knowledge DB so future local searches find them too.
• Opt-in & gated: controlled by a new citations.live_search config flag (default on wherever paper citations are configured). The tool description and EEGLAB prompt tell the agent to use it only when the user asks for recent / latest / new papers, or when the local index comes up short, so there's no surprise latency on ordinary questions.
• Responsive: the opencite call is bounded by a timeout (default 20s); on timeout/error the tool returns a graceful message and the chat continues.

Design

• search_papers_live() in papers_sync.py reuses the existing opencite client, the _run async bridge (safe under LangGraph's async ToolNode, which runs sync tools in a worker thread), _build_config, and _store_papers. API keys are read from the server's env vars (same ones OSA settings use).
• Discovery-only, consistent with the local paper tool: results are presented as references, not used to formulate answers.

Tests

• _paper_to_result mapping (real Paper, offline).
• Live network test: returns recent papers and verifies they're cached into a real SQLite DB (no mocks).
• Tool registration: present with include_live_papers=True, absent by default.
• All 6 community configs still validate with the new flag; 513 tests pass across tools/knowledge/scheduler; ruff clean.

Verified locally

Live call returned 2026 papers, newest-first, e.g. for "EEGLAB independent component analysis".

[neuromechanist]

Review (Sonnet, pr-review-toolkit) + resolutions

No critical issues. The reviewer confirmed the async/timeout reasoning is sound: __aexit__ does close all httpx clients on timeout (no leak); _run's thread-offload path runs wait_for correctly; no exception escapes to the agent; gating logic is None-safe; _paper_to_result empty-title risk is negligible (opencite filters); tests are real (no behavioral mocks).

Both 'important' findings addressed (no tech debt carried forward):

1. Cache write could block the chat response up to ~5s if the scheduler was mid-sync on the same DB. Caching now runs in a background daemon thread (_cache_papers_async) with error logging, so it never adds latency to (or can fail) the response. Tested deterministically offline via the returned joinable thread.
2. Orphaned per-source tasks on the timeout path logged spurious errors. Mitigated downstream by setting opencite's per-request config.timeout just under the overall wait_for cap so sources finish/time-out cleanly first. The proper fix belongs in opencite (cancel its own tasks on cancellation) — filed as search(): cancel in-flight per-source tasks on cancellation/timeout neuromechanist/opencite#40.

26 offline + 1 live network test pass; ruff clean.