v1.20.4

Internal — Phase 1 of the architecture cleanup

Foundational refactor with no user-visible behaviour change. First step
of the 5-phase plan in docs/markdown-text-architecture-plan.md.

• New module src/text_layout.rs — single source of truth for
  display-width calculation over ratatui span lists.
  • WrappedSpan { content: String, style, width: u16 } — owned styled
    chunk with cached display width.
  • WrappedLine { spans, width } — one wrapped visual row.
  • wrap_spans(spans, max_width) -> Vec<WrappedLine> — greedy
    word-wrap; algorithm ported verbatim from
    table_modal::wrap_cell_spans so a Phase 2 swap is mechanical.
  • measure(spans) -> u16 — total display width without allocation.
• visual_rows::line_visual_rows is now a 4-line adapter over
  wrap_spans. The old hand-written ceil-div on UnicodeWidthStr
  is gone; layout-width math has one implementation.
• state::current_line_width and highlight::apply_block_highlight
  use text_layout::measure instead of inline span-width sums.

Tests: +14 cases in text_layout::tests, including a width-sweep
harness over [20, 40, 60, 80, 120, 200], idempotence (soft-wrap
inputs only — explicitly documented), hard-newline consumption,
combining-mark glue, wide CJK, mixed styles across wrap boundaries,
and max_width == 0 short-circuit. 266 binary tests + 351 mermaid-text
tests still pass; clippy + fmt clean.

Quality gates audited (per docs/markdown-text-architecture-plan.md):
no dead code, no #[allow(dead_code)], no unused dependencies, no
duplicated width-sum loops anywhere outside text_layout::measure,
rustdoc on every pub item.

Phases 2 + 3 (wrapped-cell tables, deletion of visual_rows.rs once
prose owns its wrapping) build directly on this module.