Summary

Tab acceptance scanned to the next whitespace boundary to find "a word", so in scripts written
without spaces between words (Chinese, Japanese, Thai, Lao, Khmer, ...) a single Tab swallowed the
entire suggestion run instead of advancing one word. When an acceptance token begins with such a
script, it is now segmented with ICU word breaking and only the first segment is accepted, matching
the one-word-per-Tab feel of space-delimited text. Space-delimited scripts never enter the new
branch, so Latin / Cyrillic / Arabic acceptance is byte-for-byte unchanged.

Validation

xcodebuild ... test ... CODE_SIGNING_ALLOWED=NO CODE_SIGNING_REQUIRED=NO \
  -only-testing:CotabbyTests/SuggestionSessionReconcilerTests \
  -only-testing:CotabbyTests/SuggestionCoordinatorAcceptanceTests
# ** TEST SUCCEEDED ** — 72 reconciler + 6 acceptance tests, 0 failures
#   incl. a Latin-unchanged regression guard and CJK / Japanese / Thai segmentation cases

swiftlint lint --quiet   # exit 0 (clean)

Linked issues

None. Closes a confirmed gap where one Tab accepted an entire space-less-script suggestion.

Risk / rollout notes

• Latin/space-delimited acceptance is unchanged — the whitespace-run path is untouched; the new
  ICU branch only runs when a token begins with a space-less script (detected by leading scalar).
• Segmentation granularity follows the OS ICU word breaker (per-character or dictionary-based
  depending on the script/OS); tests assert the robust property (one segment, not the whole run)
  rather than a pinned word, so they don't flake on ICU differences.
• Full-suggestion acceptance (the full-accept key) still takes everything at once.

Greptile Summary

This PR fixes single-Tab over-acceptance in space-less scripts (Chinese, Japanese, Thai, Khmer, etc.) by routing the acceptance chunk through ICU's enumerateSubstrings(.byWords) whenever the leading character of a non-whitespace token is detected as belonging to a space-less script. Latin/Cyrillic/Arabic paths are untouched.

• beginsSpacelessScriptWord on Character performs a cheap scalar-value range check to gate the new branch, keeping the common space-delimited path allocation-free.
• firstSegmentedWordEnd wraps enumerateSubstrings with an early-exit and min(..., limit) clamp, correctly bounding the result to the whitespace-scanned token.
• Six new unit tests cover a Latin regression guard, CJK/Japanese/Thai run segmentation, cross-space boundary containment, and leading-whitespace preservation, with intentionally OS-agnostic "less than whole run" assertions.

Confidence Score: 4/5

Safe to merge — the new ICU branch is correctly gated and cannot affect Latin/Cyrillic/Arabic acceptance paths.

The core logic is correct and well-tested. The only gap is that CJK Unified Ideograph Extensions C, D, E, and F are missing from the range table, so rare archaic Han characters starting a token would still trigger the old whole-run acceptance. The impact is narrow and the fallback is the previous behavior rather than a crash or data loss.

The beginsSpacelessScriptWord switch in SuggestionSessionReconciler.swift needs the four missing CJK extension ranges added.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[nextAcceptanceChunk called] --> B[Skip leading whitespace to tokenStart]
    B --> C[Scan forward to next whitespace to index]
    C --> D{tokenStart < index AND leading char is space-less script?}
    D -- No --> F{autoAcceptTrailingPunctuation = false?}
    D -- Yes --> E[firstSegmentedWordEnd: enumerateSubstrings byWords stop after first segment]
    E --> E2{wordEnd found and wordEnd > tokenStart?}
    E2 -- Yes --> E3[index = min wordEnd limit]
    E2 -- No --> F
    E3 --> F
    F -- Yes --> G[wordEndTrimmingTrailingPunctuation peel trailing non-word chars]
    G --> H[Return remainingText up to index]
    F -- No --> H

_{Reviews (1): Last reviewed commit: "Accept one word at a time in space-less ..." | Re-trigger Greptile}

Greptile also left 2 inline comments on this PR.