feat(parser): two-phase block scanner (Phase 1)

[luca-chen198]

Summary

Replaces the single-pass regex tokenizer with a two-phase parser: a new BlockScanner walks the source line-by-line and emits typed BlockSpans for block-level constructs (paragraph, ATX heading, fenced code, thematic break, link reference definition). The existing inline-regex pass then runs only over the content of blocks that allow inline formatting, and any inline match landing inside a non-inline-allowing block is dropped at the end.

This is largely foundation work — the user-visible surface barely changes today. The motivation is to unblock Phase 2 (lists, blockquotes, tables, footnotes as proper container blocks) and to fix a class of overlap bugs structurally rather than via spreading heuristic checks.

What changes

Parser

• New Sources/MarkdownEngine/Parser/BlockScanner.swift — line-walker classifying paragraph / ATX heading / fenced code / thematic break / link reference definition.
• New Sources/MarkdownEngine/Parser/BlockSpan.swift — typed BlockKind enum with Phase-1 cases implemented and Phase-2 cases forward-declared (`.blockquote`, `.table`, `.tableRow`, `.tableCell`, `.list`, `.listItem`, `.footnoteDefinition`, `.definitionList`, `.htmlBlock`).
• New `Sources/MarkdownEngine/Parser/BlockVisitor.swift` — depth-first visitor protocol for block trees (default `walk(_:)` already recurses into children so Phase-2 nested containers work without changes to callers).
• `MarkdownTokenizer.parseTokens` now calls `BlockScanner.scan` first, runs the existing inline regexes, then filters inline tokens by block precedence — anything matching inside a fenced code block, thematic break, or link-reference-definition is dropped.

What's user-visible

• LaTeX, wiki-link, image-embed, inline code, and emphasis regex matches inside fenced code blocks are now correctly suppressed (previously they could leak through under specific patterns — `$$x^2$$` or `[[link]]` inside triple-backtick fences could render as real LaTeX/links).
• ATX heading marker space (e.g. the space between `#` and the title) now collapses cleanly when the heading is inactive — was a regression introduced by an earlier Phase-1 commit, fixed in `d1d8c02`.
• No new public API surface.

What's deliberately NOT in this PR (Phase 2)

The following `BlockKind` cases are forward-declared but `BlockScanner` does not yet emit them — anyone building Phase 2 can flip them on without an API break:

• `blockquote` — `> quoted text`
• `list(ordered:)` / `listItem(indentColumns:)` — bullet / numbered lists as container blocks (task checkboxes still work via the separate regex pass in the styler)
• `table` / `tableRow` / `tableCell(alignment:)` — pipe tables
• `footnoteDefinition(label:)` — `[^1]: footnote text`
• `definitionList`, `htmlBlock`

Thematic breaks and link reference definitions ARE detected by the block scanner but have no styler / no resolver yet — they render as their literal source text, identical to behaviour on `main`. The detection itself is groundwork for Phase 2.

What was tried then removed during review

Setext headings (`Title\n====` / `Title\n----`) were initially supported then removed. CommonMark §4.3 requires the underline to absorb all preceding paragraph lines into the heading, which is unintuitive for casual notes: a user typing `---` as a visual separator unexpectedly promotes the prior N lines into an H2. Bear, Apple Notes and Notion don't ship Setext for the same UX reason. ATX (`# Title`) covers the same case unambiguously.

Trade-offs (honest)

• ~1500 net lines of new code for foundation that is mostly invisible at the user level today.
• The visible payoff is small unless Phase 2 follows — primarily the fenced-code-overlap bug class and the heading-marker cleanup.
• Test coverage was added during development and removed at the end of the review pass. The pre-existing public-API contract suite (`MarkdownEngineDecouplingTests`, 10 tests) stays. Phase 2 work should re-add tests for whatever it builds on top.

Test plan

• `swift build` clean.
• `swift test` — 10 tests in `MarkdownEngineDecouplingTests` pass.
• Manual in an editor:
  • `# Heading` — confirm the space between `#` and the title shrinks together with the `#` when the cursor leaves the heading.
  • Paste a fenced code block containing `$$x^2$$` and `[[fake link]]` — both should render as plain code (no LaTeX render, no wiki-link styling).
  • `Title\n----` — confirm this is paragraph + literal `----`, not an H2 (Setext is deliberately disabled).
  • All existing markdown features (bold, italic, wiki links, images, LaTeX in paragraphs, task checkboxes, ATX headings of all levels, fenced code with language tag) render exactly as before.

🤖 Generated with Claude Code