feat(papers): local-first live search, confirmed + faster

[neuromechanist]

Summary

Addresses dogfooding feedback on the live paper search (#308): it felt slow and got invoked too eagerly, leaving the user waiting with no warning. Closes #312.

Changes

1. Local-first + explicit confirmation. The tool description and EEGLAB prompt now require the agent to search the local indexed library first, then offer a live search and wait for the user to confirm before running it. It must never call the live tool as the first action on a paper question.
2. Announce the wait. When the user confirms, the agent says "OK, let me search through the latest literature - this might take a minute." before invoking it.
3. Much faster. Live search now queries OpenAlex only (LIVE_SOURCES) instead of also waiting on Semantic Scholar (~1 req/s) and PubMed. Measured ~0.6s vs the previous multi-source wait. Default timeout tightened to 15s. Batch sync still uses all three sources for coverage.

Interaction (intended)

User: what are the latest papers on ICLabel?
Assistant: [calls search_eeglab_papers] Here's what's in our library: ... I can also search the latest literature - it takes about a minute. Want me to?
User: yes
Assistant: OK, let me search through the latest literature - this might take a minute. [calls search_eeglab_papers_live]

Tests

• search_papers_live gains a sources arg (defaults to LIVE_SOURCES); timed live call returns 2026 papers in ~0.6s.
• 242 tool/knowledge tests pass (incl. live network + cache); EEGLAB config validates with the new prompt; ruff clean.

Builds on #310.

[neuromechanist]

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

No critical issues. Confirmed solid: LIVE_SOURCES is an immutable tuple (safe default); sources threads through correctly; sync_all_papers still uses all 3 DEFAULT_SOURCES (no cross-contamination); timeout arithmetic fine; sort="year" recency preserved with OpenAlex-only; YAML structure valid with no leftover 'call live directly' examples; existing live/cache tests remain valid.

One 'important' finding addressed: the tool description's 'before calling it, tell the user...' implied pre-call sequencing a tool description can't enforce. Reworded to 'when you call it, your message in that turn should first tell the user...', matching the YAML framing - matters for communities that rely on the description without extra prompt guidance.

32 tool tests pass; ruff clean.

[neuromechanist]

Full PR-review-toolkit pass (Sonnet) - all findings addressed

Ran code-reviewer + silent-failure-hunter + pr-test-analyzer + comment-analyzer.

silent-failure-hunter

• Broad except Exception hid config + programming errors as empty results. Now: APIKeyError/ConfigurationError -> logger.error + []; OpenCiteError (transient/rate-limit) -> warning + []; other exceptions propagate instead of looking like 'no papers'.
• Cache-write failure logged too quietly. Escalated to logger.error(..., exc_info=True) - a lost cache write is a real degraded state.
• Tool message falsely implied a timeout for genuine empties. Reworded.

comment-analyzer

• 'up to a minute' overstated the 15s cap ~4x. Corrected to '~15 seconds' (tool description) / 'a few seconds' (prompt). Note: this softens the literal 'a minute' wording from the original ask, since live search is now ~1-5s; happy to revert if preferred.
• Tightened DEFAULT_SOURCES/LIVE_SOURCES rationale comments (peer-reviewed focus; server-side recency sort).

pr-test-analyzer

• Added TestSourceConstants: live = OpenAlex-only and a strict subset of batch DEFAULT_SOURCES (guards the core contract deterministically).
• Live test now passes sources explicitly + the production 15s timeout.

code-reviewer (earlier): description sequencing wording aligned with prompt.

31 offline + 1 live + 33 tool/config tests pass; ruff clean.