refactor(epub): Japanese CJK line-breaking PR E — WordJoin setup

[mtskf]

Japanese / CJK Line-Breaking — PR E (Phase 2: WordJoin setup)

Second PR of the Japanese CJK line-breaking effort (follows PR D, Phase 1 scaffolding). This is a
SETUP PR: it adds the UTF-8 predicates, the 3-state WordJoin model, the addWord(WordJoin) API,
the parser bridge rename, and layout-level Japanese tests. The cluster assembler is NOT wired into
the parser yet (that is PR F), so the parser still chunks Japanese exactly as today.

Behavior scope (read this)

• English / Latin layout: UNCHANGED (all Phase 1 characterization tests stay green).
• Common Japanese text wrapping: UNCHANGED (the parser produces the same chunks as master; the new
  WordJoin::CjkBreak state exists but is never produced at runtime in this PR).
• Two intentional, narrowly-scoped CJK setup changes land here (front-loaded so PR F is a clean
  assembler-wiring diff), each affecting only rare/edge inputs:
  1. utf8IsCjkBreakable is extended to supplementary CJK ranges. Its pre-existing consumer
     Hyphenator therefore suppresses inserted hyphens for those supplementary-plane CJK
     characters too — a consistent extension of the existing "CJK gets no hyphen" rule, observable
     only for text containing supplementary-plane CJK.
  2. With focus-reading enabled, a single CJK grapheme cluster is no longer focus-bolded (a
     boundary-isolated CJK char that master would have bolded now renders regular). This is the
     correct end-state for per-character CJK tokens (PR F) and is unit-tested here.

Why

The layout join flag was a single bool (wordContinues) with only two states — "space +
breakable" (word gap) and "no-space + non-breakable" (no-break-space glue) — but CJK needs a third:
"no-space + breakable" (break between ideographs, no space). This PR generalises the flag to a
3-state enum and makes CJK codepoints first-class breakable units at the predicate level. Every
read site is routed through needsSpaceBefore (space-vs-kerning gap) / canBreakBefore (break
allowed). The two legacy states map identically (false⇔Space, true⇔Glue); the predicates
differ only on CjkBreak, so Latin/legacy output is bit-identical.

Commits

1. feat(utf8): extend utf8IsCjkBreakable and add utf8IsGraphemeExtender
   • utf8IsCjkBreakable extended to supplementary CJK ranges (Kana Supplement/Extended-A/B, Small
     Kana Extension, CJK Extensions D/E/F/G/H/I/J, Compatibility Ideographs Supplement).
   • New utf8IsGraphemeExtender (combining marks, kana voicing U+3099-309A / U+FF9E-FF9F,
     ideographic tone marks U+302A-302F, variation selectors U+FE00-FE0F / U+E0100-E01EF).
   • Isolated host gtest target test/utf8/ (links only Utf8.cpp).
2. refactor: model word joins as 3-state WordJoin (space/glue/cjk-break)
   • New lib/Epub/Epub/WordJoin.h (enum class WordJoin {Space,Glue,CjkBreak} + needsSpaceBefore
     / canBreakBefore constexpr predicates).
   • ParsedText.{h,cpp}: wordContinues bool → WordJoin; every read site routed through the
     predicates; isSingleCjkGraphemeCluster focus-bypass helper (NFC + NFD).
   • Parser bridge: nextWordContinues → nextJoin (every =true→Glue, =false→Space),
     in the same commit as the addWord signature change so the firmware compiles mid-bisect.
3. test: Japanese text wraps at viewport width, not at fixed chunk
   • Layout-level Japanese tests: viewport-wrap (not fixed-chunk), Latin-survival in a CJK run,
     Space/Glue/CjkBreak join transitions, focus-bypass for NFC + NFD CJK clusters.
4. fix(epub): preserve underline on CJK focus-bypass and guard embedded NUL
   • Cross-model (Codex) review follow-up: the focus-bypass now pushes the underline-applied style
     (was dropping underline on a single underlined CJK cluster), and isSingleCjkGraphemeCluster
     is length-bounded (rejects an embedded NUL instead of truncating validation). Adds a regression
     test pinning underline survival through the bypass.

Out of scope (deferred to PR F)

• Utf8ClusterAssembler and its wiring into ChapterHtmlSlimParser (parser still chunks as today).
• SECTION_FILE_VERSION bump (no cache-format change).
• Kinsoku (禁則処理), vertical text, GfxRenderer cluster-aware rendering.

Known pre-existing limitation (follow-up, NOT introduced here)

The focus-reading tokenizer in ParsedText::addWord (the while (ptr < end) loops at
ParsedText.cpp:220 and :270) can infinite-loop on a word containing an embedded NUL byte, because
utf8NextCodepoint returns 0 without advancing the pointer at a NUL. This hazard exists on master
independently of this PR (a Codex review surfaced it). This PR does not introduce or regress it:
isSingleCjkGraphemeCluster's NUL guard (commit 4) returns the CJK-base+NUL case to master's
behavior. The input is practically unreachable from EPUB content (XML 1.0 forbids U+0000; expat
rejects it), so it is deferred to a dedicated focus-tokenizer robustness PR rather than expanding
this behavior-unchanged setup PR.

Verification

• Host cmake gtest: full suite 117/117 PASS, including the 4 Phase-1 ParsedTextLayout cases
  UNCHANGED (behavior-preservation oracle) + the new Japanese cases + test/utf8 predicates +
  the underline regression. Focus-bypass non-vacuity confirmed by revert-check.
• Cross-model review: Codex pass over the full diff — read-site mappings, scope guards, and
  no-runtime-CjkBreak all confirmed; the two findings it raised (underline drop, embedded NUL) are
  fixed in commit 4.
• Firmware build verified by CI (no PlatformIO on the dev machine); clang-format applied by CI.