Capture workflow-pass conventions in subagent rules

The ranking/dedup/filter/concurrency workflow pass introduced two conventions
worth pinning so they don't bit-rot:

- code-quality-reviewer (Async & Concurrency): add the bounded-fan-out rule —
  a stage that gathers one op over N papers (enrich / download / OA resolve)
  escapes the per-source token buckets and must use a module-level Semaphore
  cap. Includes the 25-paper 429 / IP-block rationale + anti-pattern + pattern.

- compliance-auditor (Core vs Source Plugins): add the query-semantics rule —
  Query filter fields (min_citations / year / top_tier) are enforced in the
  pipeline for all sources, never delegated to plugins. Documents the real
  min_citations incident (only semantic_scholar honoured it) + anti-pattern.

Author: JE-Chen

.claude/agents/rules/code-quality-reviewer.md

@@ -51,6 +51,7 @@ You do NOT modify files. The parent agent decides whether to fix.
 - Per-source rate limits live in `thesisagents/fetchers/rate_limit.py` as a token-bucket decorator. Each source plugin declares its own bucket (`arxiv: 1 req/3s`, `semantic_scholar: 1 req/s`, `scholar: 1 req/10s with jitter`, etc.). Do NOT bypass the bucket — even retries go through it.
 - Streamlit runs the UI on a separate thread per session. Mutate `st.session_state` only, never module globals. Long-running export jobs are dispatched to the FastAPI backend and polled.
 - All fixture-recording, CLI exports, and tests use `asyncio.run` at the outermost layer and never inside library code.
+- **Bounded fan-out over a paper collection.** A per-source token bucket caps traffic *within* one source, but a stage that fans one operation out over N papers (PDF download, OA resolve, LLM enrich) escapes those buckets — several papers can hit the *same* external API (Anthropic) or publisher CDN (`dl.acm.org`) at once through different source clients. Any `asyncio.gather(*(coro(item) for item in collection.papers))` MUST wrap each task in a module-level `asyncio.Semaphore` cap (the project uses 4–5: `pipeline._ENRICH_CONCURRENCY`, `pdf_download._DOWNLOAD_CONCURRENCY`, `oa_resolver._CONCURRENCY`). **Why:** an unbounded fan-out of a 25-paper search fires 25 concurrent Anthropic calls (an easy 429) or 25 hits on one CDN (an IP block) — it looks fine at 3 papers and fails at 30. **Anti-pattern:** `await asyncio.gather(*(_download_one(p, d) for p in collection.papers))` with no semaphore. **Pattern:** define `sem = asyncio.Semaphore(max(1, concurrency))` and a `_bounded` wrapper that does `async with sem: return await coro(item)`.
 
 ## Security (review-level)
 

.claude/agents/rules/compliance-auditor.md

@@ -46,6 +46,8 @@ The line between `thesisagents/` and `sources/<name>/` is **not** "anything sour
 
 **Loading source plugins:** at runtime `thesisagents/fetchers/base.py::load_fetcher(name)` does `importlib.import_module(f"thesisagents.sources.{name}")`. No more `sys.path` injection — sources are a regular sub-package of `thesisagents`. Plugins under `thesisagents/sources/<name>/` use **relative imports** internally (`from .fetcher import …`, `from .parser import …`, `from . import webrunner_backend`). Tests import them via the full path (`from thesisagents.sources.<name>.fetcher import …`) and reference module-string targets the same way (e.g. `monkeypatch.setattr("thesisagents.sources.acm.fetcher.get_client", …)`).
 
+**Query semantics are enforced in core, never delegated to plugins.** Every filter field on `Query` (`min_citations`, `year_from` / `year_to`, `top_tier_only`) MUST be applied by the pipeline (`thesisagents/core/pipeline.py::run_search`, after dedup + rank) so it holds for **all** sources uniformly. A plugin MAY *additionally* push the filter to its upstream API as an optimisation (e.g. `semantic_scholar` sends `minCitationCount`), but core must never *rely* on plugins to do it. **Why (real incident):** `min_citations` was accepted by CLI/MCP/`Query`, documented as a threshold filter, and implemented by only the `semantic_scholar` plugin — so for the other 14 sources it silently did nothing; `min_citations=50` still returned 2-citation arXiv / OpenAlex / DBLP papers. **Anti-pattern:** wiring a `Query` filter into one source's request params and assuming the result set is filtered. **Pattern:** filter in `run_search`, and keep records whose value is *unknown* (`citation_count is None` / `year is None`) rather than treating unknown as failing the threshold (a source not reporting a count must not silently drop the paper).
+
 **When in doubt:** "if a user installs ThesisAgents with the default `requirements.txt` and never enables a source plugin, should this source work?" Yes → core. No → source plugin.
 
 ---