fix(acceptance): CJK phrase boundaries and decode stop, so Japanese accepts advance clause by clause

[FuJacob]

Summary

A Japanese phrase accept could arrive as one giant chunk (reported tail: 理解し、その内容を自分の言葉で表現する。 in a single Tab) because every layer only knew ASCII punctuation: phrase mode never stopped at 。 or 、, a punctuation-led chunk skipped ICU segmentation and swallowed everything to the next whitespace, and SentenceBoundaryClassifier never registered a CJK sentence end so the decode stop policy let generations run to the full token budget. Phrase boundaries now include the CJK terminators (。！？｡) and clause commas (、，､), word chunks bind trailing CJK punctuation to their word (読み、 is one Tab), punctuation-led runs peel as their own chunk (including opening brackets, so flat quoted runs like 「分かった」と言った no longer swallow whole), and generation stops at the end of a Japanese sentence like it does for English. All added codepoints occur only in CJK text and ASCII ,/brackets stay non-boundaries, so space-delimited scripts are byte-for-byte unchanged.

Validation

• xcodebuild build-for-testing ... CODE_SIGNING_ALLOWED=NO then test-without-building for the full unit bundle: ** TEST SUCCEEDED **, 993 tests, 0 failures (4 pre-existing skips). New cases cover the reported clause (理解し、 stops at the comma), 。/！/？ stops, the 終わり。」 closer walk-back, the leading-punctuation cliff, the opening-bracket peel (「分かった」と言った → 「), mixed closer-opener runs, halfwidth ､ ｡ ｣ parity, auto-accept-off re-peeling, and ASCII comma/bracket non-boundaries for English.
• swiftlint lint --quiet on all changed files: exit 0, no findings.
• Not yet validated end-to-end on device with a live Japanese session; the chunking rules are pure Support/ logic locked by the unit tests above.

Linked issues

None filed; follow-up to the Japanese IME report behind #668 (same user feedback thread, separate root cause).

Risk / rollout notes

• Behavior change for CJK phrase-granularity users: Tab now advances clause by clause (、) and stops at 。, instead of accepting an entire space-less sentence per press. Word granularity gains the punctuation binding (資料、 is one chunk), loses the punctuation cliff, and peels opening brackets as their own chunk.
• Generation for CJK text now early-stops at sentence ends via the existing decode stop policy, so Japanese/Chinese suggestions get shorter and cheaper at the source; English generation is unchanged.
• Same pre-existing CI caveat as fix(insertion): IME-safe accept so Japanese/CJK suggestions land on Tab #668: main requires a cotabbyinference revision that may not be resolvable by CI until the package push lands.

Greptile Summary

This PR fixes CJK (Japanese/Chinese) text acceptance in phrase mode by threading phrase-boundary awareness through three layers: the decode stop policy, the phrase accumulator, and the word-chunk extractor. Before this change, a Japanese phrase could arrive as one giant Tab because every layer only knew ASCII punctuation.

• SentenceBoundaryClassifier gains CJK sentence terminators (。！？｡) so the decode stop policy fires at the end of a Japanese sentence instead of running to the token budget, and the closing-punctuation walk-back now skips CJK brackets so 終わり。」 registers as a sentence end.
• SuggestionSessionReconciler adds endOfCJKPunctuationRun to bind trailing 、 to the preceding word (読み、 is one chunk), a punctuation-led peel so opening brackets like 「 don't swallow flat quoted runs, and endsAtPhraseBoundary replaces endsInSentenceTerminator to stop the phrase accumulator at ideographic commas as well as terminators.
• CJK primitives are consolidated into a single Character extension so a future codepoint addition updates all three policies in one edit.

Confidence Score: 5/5

Safe to merge; all changes are additive, CJK-only string-parsing logic with no impact on ASCII/space-delimited text, validated by 993 passing unit tests.

All changes are additive CJK-only string-parsing logic, ASCII/space-delimited paths are byte-for-byte unchanged and covered by existing tests, and every new code path has dedicated unit tests including edge cases and policy-flag interactions.

No files require special attention beyond the minor doc-comment omission in SuggestionEngineModels.swift.

Important Files Changed

Filename	Overview
Cotabby/Support/SuggestionSessionReconciler.swift	Core chunking and phrase-accumulation logic updated; adds endOfCJKPunctuationRun, trailing-binding, punctuation-led peel, and endsAtPhraseBoundary. Interaction between CJK binding and autoAcceptTrailingPunctuation=false is intentional and now tested.
Cotabby/Support/SentenceBoundaryClassifier.swift	Adds CJK sentence terminators to the decode-stop switch and extends isSentenceClosingPunctuation to include CJK closers; both changes are additive and well-scoped, with full test coverage.
Cotabby/Models/SuggestionEngineModels.swift	Doc-comment update only; no logic change.
CotabbyTests/SentenceBoundaryClassifierTests.swift	Adds 5 new tests for CJK terminators, closer walk-back, halfwidth parity, and ideographic comma non-termination; all appear correct.
CotabbyTests/SuggestionSessionReconcilerTests.swift	Adds 17 new tests covering all new chunking rules, phrase-boundary stops, halfwidth parity, ASCII non-regression, and autoAcceptTrailingPunctuation interactions; coverage looks thorough.

_{Reviews (3): Last reviewed commit: "Address Greptile review on #669: single-..." | Re-trigger Greptile}

[FuJacob]

Addressed the review finding: the CJK terminator and closer sets are now declared once (internal Character extension) and composed by both the phrase policy and SentenceBoundaryClassifier, so a future codepoint is a single edit. Local validation on the rebased branch (top of merged #668): xcodegen no drift, swiftlint clean, build-for-testing succeeded, full unit bundle 1001 tests / 0 failures.