Add source-type aware extraction and avoid binary/text fallbacks in extracted.md

[jfraser]

Summary

wiki_capture_source would be more robust with a source-type aware extraction pipeline. Today, fallback behavior can write non-markdown content into extracted.md, including raw PDF bytes or raw XML.

Problem

The packet contract says:

raw/sources/SRC-*/extracted.md — normalized markdown text

But current behavior can produce:

• raw PDF bytes if MarkItDown times out and curl fallback succeeds,
• raw XML for local .xml files,
• potentially raw HTML or other text-like formats without normalization.

Suggested design

Use a typed extraction pipeline:

1. Identify source type:
   • URL extension
   • Content-Type
   • file extension
   • magic bytes
2. Route to appropriate extractor:
   • PDF → download original, run MarkItDown/PDF extractor
   • HTML → readability/markdown extraction
   • XML → XML-to-markdown or project-provided converter
   • Markdown/text → copy as-is
   • binary/unknown → write clear extraction failure message
3. Record extraction metadata in manifest:
   • extractor
   • extraction_status: success | failed | timeout | unsupported
   • content_type
   • original_file
   • optional error message

Expected behavior

extracted.md should always be human-readable markdown/text, or a clear extraction failure note. It should never contain binary bytes.

Optional extension point

A pluggable extractor interface would let projects register domain-specific converters for structured XML or other specialized source formats.