feat: add CRWSearchRM (fastCRW) retrieval module

[us]

What

Adds CRWSearchRM, a new retrieval module backed by fastCRW, alongside the existing RMs (Tavily, Serper, Bing, DuckDuckGo, SearXNG, …). Selectable with --retriever crw.

Why

The README explicitly invites new-retriever PRs, so here's one.

What makes fastCRW worth adding (not "yet another Firecrawl")

1. The only self-hostable engine that actually works on protected/JS-heavy pages.
Firecrawl's OSS build gates its real anti-bot layer (fire-engine) behind a cloud-only flag — self-hosted Firecrawl falls back to plain fetch or vanilla Playwright and fails on Cloudflare-protected or heavily JS-rendered pages. fastCRW ships the complete stack in a single ~8 MB Rust binary: Cloudflare JS-challenge handling, UA rotation + BYO-proxy support, SPA rendering, and an HTTP→headless→proxy fallback ladder — all in the open core (AGPL), no cloud dependency, no closed components.

2. Faster and higher-recall on Firecrawl's own benchmark dataset.
On the public Firecrawl benchmark, fastCRW scores 63.74% truth-recall vs 56.04% for Firecrawl, with a faster median latency (p50 ~1.9 s vs ~2.3 s). Single binary, ~6 MB RSS — no Redis, no worker processes, no Playwright sidecar.

3. A managed upgrade path for the SearXNG RM.
On web search specifically: crw is not an alternative to SearXNG — it is built on top of it. SearXNG is the metasearch aggregator underneath; crw adds a quality layer above it: query expansion (multi-variant rewrite), content-aware reranking (re-scoring by fetched content rather than SearXNG's content-blind ordering), and category routing (research queries fan out to arXiv / Semantic Scholar / Google Scholar, code queries to GitHub). So you get SearXNG's breadth plus a measurable accuracy layer, all open-source (AGPL) and self-hostable — or just point at the managed endpoint without standing up your own instance.

The integration diff is small because fastCRW implements the Firecrawl API surface, making it a drop-in — but that compatibility is the reason the patch is tiny, not the reason to use fastCRW.

Changes (additive only — no existing RM touched)

• knowledge_storm/rm.py: CRWSearchRM, mirroring TavilySearchRM (same constructor, same forward()/get_more_info return shape and usage accounting).
• tests/test_crw_rm.py: unit tests with mocked HTTP (no network).
• Wired --retriever crw into every examples/ runner + their docstrings; one-line README note.

Config

export CRW_API_KEY=...   # https://fastcrw.com/dashboard (free tier available)
# optional: point at a self-hosted server instead of the managed cloud

Happy to adjust anything to match your conventions. I maintain the integration and am glad to set the team up with free credits to evaluate it.