Keep Chinese phrases together in HTML reports

Traceability

• Spec ID: 2026-07-30-html-cjk-line-breaking
• Status: Implemented
• Issue: None; focused follow-up to the reviewed HTML progressbar visual QA.

Intent

Keep ordinary Chinese word-like phrases together when the self-contained HTML report wraps headings and card copy. Segment visible zh-CN text with the standards-based Intl.Segmenter API and emit bounded, nonvisual keep-together spans so the result is deterministic across supported browsers. The generic renderer must improve reviewed Chinese content without hardcoded phrases, report-data mutation, runtime measurement, or a visual redesign.

Acceptance Scenarios

• AC-1: A zh-CN HTML report uses Intl.Segmenter word segmentation for visible text and wraps each ordinary all-Han word-like segment of at most eight code points in <span class="cjk-phrase">. Representative phrases including 口径, 入口, 结论, and 命令 receive that deterministic markup.
• AC-2: Phrase markup remains scoped to visible Chinese report text. English reports retain their current markup, wrapping behavior, and visible copy; attributes, accessibility labels, copied prompts, and embedded report JSON remain plain escaped text without phrase spans or invisible joiners.
• AC-3: Chinese headings, card titles, summaries, reasons, and ordinary paragraphs continue to wrap between marked segments. Each keep-together span is bounded to eight Han code points. Long tokens, paths, URLs, mixed Latin text, and unrecognized segments remain escaped readable text and cannot gain an unbounded no-wrap container.
• AC-4: The existing dimension progressbar role, label, range, rounded value, count validation, CSS, score computation, report data, and artifact set remain unchanged.
• AC-5: Focused HTML tests, the full package check, whitespace validation, and the allowed-file audit pass. Before review completion, an environment with a browser runtime inspects the same report at all three target widths.

Compatibility Boundary

Intl.Segmenter is used only while generating the self-contained report. The renderer accepts only isWordLike segments that consist entirely of Han code points and are between two and eight code points long. All other segments use the existing HTML escaping path. If Intl.Segmenter is absent or throws, the entire value falls back to the same escaped readable text. CSS applies white-space: nowrap only to the bounded phrase spans, leaving normal wrapping available between spans and around Latin tokens, punctuation, paths, and URLs.

Non-goals

• Hardcoding known report phrases or inserting word-joiner characters into report data, visible copy, accessibility text, or copied prompts.
• Preventing ordinary Chinese paragraphs from wrapping or forcing a complete sentence, heading, or card onto one line.
• Adding overflow, clipping, ellipsis, hidden content, fixed widths, custom fonts, browser-specific JavaScript, dependencies, or generated artifacts.
• Changing English layout, scores, findings, dimensions, report schemas, progressbar semantics, interactions, or artifact names.

Plan and Tasks

1. Add a focused renderer regression for deterministic markup, escaping, bounded long-token handling, mixed Latin/path/URL preservation, English preservation, unavailable-segmenter fallback, and the four reviewed phrases. Record the missing-markup failure before renderer implementation. (AC-1..AC-4)
2. Add one visible-text renderer using Intl.Segmenter and a bounded .cjk-phrase keep-together rule. Route visible report copy through it while leaving attributes, copied prompts, and serialized data unchanged. Remove the insufficient browser-dependent auto-phrase contract. (AC-1..AC-4)
3. Update the public changelog and verify the exact reviewed Chinese report at 375, 768, and 1280 pixels, including wrapping, overflow, clipping, console, and preserved progressbar semantics. (AC-3..AC-5)

Test and Review Evidence

• The initial focused test failed for the intended reason: the reviewed Chinese phrases had no deterministic keep-together markup. It passed after the standards-based segment renderer and bounded phrase span were added.
• Browser QA then exposed descendant-selector leakage: nested phrase spans in the score orbit, metrics, and evidence grid inherited block, width, and label typography rules. A focused regression failed on the missing direct-child selector contract before those existing rules were scoped to their direct label and value children; nested phrase spans now remain inline.
• The combined CJK and preserved progressbar suite passed 2/2: node --test --test-name-pattern='HTML CJK phrase|HTML dimension progressbar' test/harness-report-render-cli.test.mjs.
• The target-owned HTML suite passed 6/6: node --test --test-name-pattern='HTML' test/harness-report-render-cli.test.mjs.
• The coordinating PM ran the complete test suite serially with an isolated temporary root: all 894 tests passed, followed by successful npm and runtime zip package verification. This resolved the parallel temporary-ancestor interference observed in two earlier Worker runs.
• The exact reviewed Chinese report retains its source scores, finding count, progressbar count, embedded report data, and artifact set. Post-fix browser review at 375, 768, and 1280 pixels remains the final visual gate.