fix(rtl): only letters vote in RTL text detection

[abimaelmartell]

Problem

is_rtl_text counts any character in the RTL Unicode blocks as an RTL vote, but the Arabic blocks embed weak-directionality characters — Arabic-Indic digits (U+0660–0669, U+06F0–06F9), number separators (U+066B/U+066C), and combining marks — that are bidi class AN/NSM per UAX #9, not strong RTL.

A line containing only Arabic-Indic digits therefore votes RTL (rtl > 0, ltr = 0) and sort_line_items reverses its item order — producing digit-reversed numbers like 0,11 instead of 11,0 — while the equivalent ASCII-digit line stays neutral.

Fix

Require is_alphabetic() for the RTL vote, mirroring the rule the LTR side already had: only letters vote in either direction. Digit-only and punctuation-only lines keep their natural left-to-right order. Divergence is confined to lines with no strong RTL letter; lines with Arabic/Hebrew letters are unaffected.

Testing

• New unit tests: Arabic-Indic digit-only lines (plain, extended, and with number separators) are not RTL; Arabic letters + Arabic-Indic digits still RTL; letters + combining marks still RTL.
• cargo fmt / cargo clippy -- -D warnings / cargo test all pass.
• Full-corpus A/B against a main-built binary: only Arabic-content documents change, restoring correct digit order in numeric fragments (e.g. a temperature-gradient table now reads 11,0 / 20,0 / 32,0 instead of digit-reversed values).

🤖 Generated with Claude Code

---

Summary by cubic

Only letters vote in RTL text detection, and we now exclude combining marks. Previously any RTL-block character (digits, separators, combining marks) voted RTL and reversed digit-only lines; digit-, punctuation-, and marks-only lines now stay neutral and keep left-to-right order, while lines with Arabic/Hebrew letters remain RTL.

• Restricts is_rtl_text to count votes only for alphabetic, non-combining chars (via unicode_normalization::char::is_combining_mark); mirrors the LTR rule and keeps the CJK exclusion.
• Adds tests for Arabic-Indic digits, separators, combining marks, and mixed-letter cases; effect is limited to sort_line_items when no strong RTL letters are present.
• Merges main (clippy 1.98 fixes and toolchain pin); no behavior change.

^{Written for commit e7a3d0f. Summary will update on new commits.}

[cubic-dev-ai]

All reported issues were addressed across 2 files

<sub>Shadow auto-approve: would not auto-approve because issues were found.

Fix all with cubic | Re-trigger cubic</sub>

[cubic-dev-ai]

0 issues found across 2 files (changes from recent commits).

<sub>Shadow auto-approve: would auto-approve. Fixes RTL text detection to count only alphabetic, non-combining characters as RTL votes, mirroring the LTR rule, so digit/punctuation-only lines stay neutral instead of reversing order; covered by new tests for Arabic-Indic digits, separators, combining marks, and mixed-letter cases.

Re-trigger cubic</sub>

[abimaelmartell]

The failing Clippy check is pre-existing toolchain drift (CI stable moved to clippy 1.98, which added lints that fire on existing code in main) — fixed separately in #442. This branch's own code passes clippy 1.98 locally.