…cle prompt

Profile of one hybrid extraction on a 50-person paper-heavy repo
(deeplabcut/deeplabcut): person stage 5:06, **article stage 12:00 on a
single LLM agent invocation that fired 100+ tool calls**, membership
stage 6:37 — total ~25 min, dominated by serial waits and an
unbounded article-agent tool-call loop.

Three independent quick wins, each tunable per deployment:

1. `max_concurrent_agents` default 3→8 in the orchestrator (and the
   `V2_MAX_CONCURRENT_AGENTS` env-resolver default in the API layer
   raised 6→8). Person/membership stages are bottlenecked on the
   `asyncio.Semaphore`, not the LLM provider — bumping the cap absorbs
   wider-fanout repos without saturating RCP.

2. New `_default_usage_limits()` in `V2LLMRuntime`: caps every
   agent invocation at 25 model requests + 50 tool calls via
   pydantic-ai's `UsageLimits`. Without a cap the article agent could
   keep cross-validating the same DOI across five tools indefinitely.
   Overridable per-call (existing kw-only signature) or globally via
   `V2_LLM_REQUEST_LIMIT` / `V2_LLM_TOOL_CALLS_LIMIT`. The cap turns
   runaway loops into clean `LLMRuntimeError` that the per-stage
   runner already handles as a per-item warning.

3. Tightened the article agent system prompt with two "stop early"
   rules: emit immediately once a concrete DOI/title is found (no
   cross-validation past two sources), and emit `{}` after two
   consecutive empty searches instead of looping. The LLM was doing
   ~25 OpenAlex calls per agent invocation chasing the same paper.

Expected impact on the profiled repo: ~25 min → ~10 min wall time. No
schema, API, or routing changes — caps are tunable and conservative
defaults.