[#39] Capture X long-form Articles as blogposts

[VGonPa]

Closes #39.

Captures X (Twitter) long-form Articles as ordered Obsidian blogposts — text and inline images interleaved where the author placed them — instead of the previous text-only trafilatura dump. Delivered as a 5-PR chain merged onto this umbrella branch, then rebased-forward onto develop.

The chain (extract → fetch/blocks → media → generate)

1. PR1 — model. New ArticleBlock tagged union (ArticleTextBlock | ArticleImageBlock) carried on ContentSourceSuccess.blocks. Additive field: a legacy record simply has no blocks key and defaults to []. Invariant: the ordered concatenation of the text-run texts equals the source's flattened text (so enrich/topics consume text unchanged).
2. PR2 — extract the Article-entity link. Resolve the article entity path in the X GraphQL payload (article.article_results.result.rest_id).
3. PR3 — fetch structured ordered blocks. Intercept the article-content GraphQL response and parse its Draft.js content_state into ordered blocks (extract.article); a text == "".join(block texts) validator + a truncation tripwire guard against a partial capture masquerading as a complete body.
4. PR4 — media: inline-image download. Each inline image wraps the existing MediaEntry photo-state union as MediaPhotoPending, so the proven xbrain media download engine, validators, and _media/ mirroring all apply with zero new plumbing.
5. PR5 — generate: blogpost render. Walk blocks in order under a ## <header>: <title> heading; text paragraphs + inline ![[_media/…]] embeds. Empty/image-only-pending bodies suppress the bare heading.

Reuse-first

Built on what already existed rather than parallel machinery: the photo download engine + MediaEntry union (PR4), _mirror_file for _media/ mirroring at generate time (PR5), the _sources_materially_equal fingerprint + injectable clock in _attach_x_sources (PR3). enrich/topics are unchanged — they still read text. The whole feature is back-compat additive: no on-wire migration, text-only fallback preserved end-to-end.

Post-review fix round (this branch, commit ac46002)

• Consolidated the duplicated "\n\n" inter-paragraph separator into a single source of truth (models.ARTICLE_PARAGRAPH_SEP), imported by both producer and consumer so they can't drift.
• Navigation-failure diagnostics in fetch_x: keep the timeout classification but capture a capped str(exc) into error + logger.warning(exc_info=True) (the article path is the still-unvalidated one).
• Aligned _fetch_tweet's non-JSON-body swallow to a DEBUG log (was a bare pass).
• Closed the untested _needs_x_fetch until-upper-bound branch.

⚠️ CAVEAT — read before relying on this in production

The X-GraphQL Article-entity path, the Draft.js content_state shape, and the intercepted GraphQL op-name are all pinned against CONSTRUCTED fixtures — they are UNVALIDATED against a real bookmarked-Article payload (RFC #39 open-Q #4).

This is safe to land: on any shape/op-name drift the parser yields no blocks and the fetch degrades cleanly to the text-only trafilatura path (never a crash, never a partial/wrong body masquerading as complete). But a real-payload validation is required before production reliance — tracked in a follow-up issue (capture one live payload, validate the entity path + content_state op-name/shape, commit it as a recorded fixture, and fold in the deferred card/unified_card variant).

Gate (green in an isolated worktree, own venv+src)

• 1053 tests passed
• ~95% coverage (gate floor 78%)
• Radon all A/B; Ruff (lint+format), Mypy, Bandit, Detect-secrets, Deptry all PASS

🤖 Generated with Claude Code