feat(capture): file a link card when a page cannot be read

[arthware-dev]

A pasted link that a site refuses to serve used to become either a silent failure or a wiki entry summarising a cookie banner. Every fetched page is now classified before anything is filed, and a blocked link still files as a link card carrying whatever the sender wrote around it.

The chat reply names the obstacle (bot check, sign-in, consent wall, paywall, no text) instead of "couldn't read that link". Recipes come from the page's own JSON-LD, so quantities are the site's rather than a summary's.

Phase 1 of docs/design/web/plan.md. Re-validation of the spike findings, including one reversal, is recorded in that doc.

[arthware-dev]

Rig verification, live against the local model (stacktests ai local, Qwen3.6-35B-A3B).

Blocked page - https://www.decathlon.de/ posted with the text "Gear list for the camping trip":

🔒 www.decathlon.de checks for bots, so I couldn't read the page. Saved the link.
✅ Captured: Camping trip gear list (Decathlon)
   camping | gear-list | decathlon | shopping | Homer

The entry is built from the sender's own words. Previously this was a silent ❌ or an entry summarising the Cloudflare challenge page.

Recipe - tier 1, JSON-LD, 968 chars, no LLM guessing at the page:

✅ Captured: Greek Salad Recipe (Griechischer Salat)
   - Servings: 4
   - Total time: 35 min
   - Ingredients: 500 g red and yellow peppers, 300 g tomatoes, 400 g cucumber, ...

Shell page - maps/place/Brandenburger+Tor/ filed as "Brandenburger Tor" via tier 0, instead of a summary of "Find local businesses, view maps".

Log lines:

[extractor] https://www.decathlon.de/ → challenge: bot protection served 'just a moment...' instead of the page
[extractor] https://www.essen-und-trinken.de/rezepte/48816-... → tier 1 (default), 968 chars
[extractor] https://www.google.com/maps/place/Brandenburger+Tor/ → tier 0 (google-maps), 71 chars

Tests: 648 framework, 2165 stacklets, both green. The 5 fixtures under tests/fixtures/web/ are real captures with inline script/style bodies stripped; provenance and the strip rules are in their README.