v0.1.4

Hardening and simplification. A day-long fuzzing campaign fixed the whole family of edge cases where reflowed output could parse differently than the input — splits manufacturing accidental tables, definitions, or headings; escapes re-pairing code spans; typography flipping raw-HTML recognition. All were pathological shapes that never occur in normal prose; each is now pinned by a regression seed. Alongside the fixes, some deliberately over-clever handling was descoped in favor of simpler, safer rules.

Fixed

• Two field reports from real docsets: hard-wrapped prose with a parenthetical spanning a line break is no longer silently skipped (#14), and a valid UTF-8 file is no longer refused when a multi-byte character straddles the 8 KB detection boundary (#15).
• Roughly twenty fuzz-found idempotency and render-preservation corners (#4–#8, #10–#13 and successors), none affecting normal documents.

Changed

• Formatting is now guaranteed stable: Format's output is a fixpoint — running the tool twice never produces a second diff, even on adversarial input.
• Invalid UTF-8 is rejected outright with the typed error mdreflow.ErrInvalidUTF8 (CLI: exit 3, even under --force); Markdown is text, and bytes with no character interpretation have nothing meaningful to reflow.
• Simpler scope around link reference definitions: any paragraph containing or directly touching a [label]:-shaped line now passes through unchanged, replacing subtle partial handling. Definitions in their own blank-line-separated block — the common layout — were already skipped, so typical documents are unaffected.
• Footnote bodies ([^label]: ...) still reflow as ordinary prose, and their continuation lines are now indented four spaces — the one spelling every Markdown renderer keeps inside the footnote.
• Paragraphs containing control characters (form feed, vertical tab, …) pass through unchanged.

Full changelog: v0.1.3...v0.1.4