[bug] web_fetch silently returns link-less Markdown when Readability.js is unavailable (Windows / npm-less environments)

[sjr666666]

Before you start

• I searched existing issues and this is not a duplicate.
• I can reproduce this on the latest main (e2feca1).

Problem summary

On hosts where Node.js is installed but npm is only resolvable as npm.cmd (stock Windows installs), readabilipy's have_npm() — which calls subprocess.run(["npm", ...]) — raises FileNotFoundError ([WinError 2]), because CreateProcess does not resolve .cmd shims. The lazy node_modules installation inside the readabilipy wheel therefore never runs, have_node() returns False on every call, and every extraction silently degrades to readabilipy's pure-Python tree.

That fallback strips element attributes, which erases the href/src destinations that _resolve_html_urls (introduced for #5307) has just resolved into the markup. The model-visible Markdown returned by web_fetch ends up with link text but no destinations.

Evidence

• Upstream root cause, already reported: Cannot find NPM in path on Windows alan-turing-institute/ReadabiliPy#115 ("Cannot find NPM in path on Windows", open since 2024-12).

• On Windows 11 (Python 3.12.13, uv sync --locked), 22 tests in backend/tests/test_web_fetch_relative_links.py fail on main (e2feca1), e.g.:

  FAILED tests/test_web_fetch_relative_links.py::test_extract_article_resolves_link_destinations[../next-https://example.com/next]
  FAILED tests/test_web_fetch_relative_links.py::test_document_base_skips_target_only_base
  ... (22 total)
  

• Minimal repro of the dependency behavior (no DeerFlow code involved):

  from readabilipy import simple_json_from_html_string
  html = ('<html><head><title>Guide</title></head><body><article><p>'
          + 'This article explains the documentation in detail. ' * 10
          + '</p><p><a href="../next">Next</a></p></article></body></html>')
  art = simple_json_from_html_string(html, use_readability=True)
  print(art["content"][-120:])
  # Warning: A working NPM installation was not found. ...
  # Warning: node executable not found, reverting to pure-Python mode. ...
  # ...</p><p>Next</p></article></body></div>   <- <a href="../next"> flattened to plain text

  shutil.which("npm") finds npm.CMD, subprocess.run(["npm", "version"]) raises FileNotFoundError, while subprocess.run(["npm.cmd", "version"]) succeeds.

Impact

• Windows users and npm-less deployments (slim containers) get web_fetch output whose links are all flattened to plain text — silently, with only a print to stderr from the dependency.
• Every extraction re-runs the availability probe, re-attempting an npm install that can never succeed in these environments (per-fetch subprocess overhead + repeated warning spam).

Proposed direction

Route unavailable-or-failing Node extraction to a link-preserving Python fallback inside deerflow/utils/readability.py (the destinations are already absolute after _resolve_html_urls; they only need to survive extraction), and cache the availability probe per process. PR incoming.