fix(core): preserve natural-language direction evidence

[CodeinScrubs]

What changed - keeps ordinary English compounds such as well-known and state-of-the-art as natural-language LTR evidence instead of discarding every hyphenated token as code - distinguishes uppercase prose (PLEASE READ THIS WARNING) from short acronyms using deterministic block context; short all-capital phrases such as HTTP API remain acronym-shaped - preserves known technical tokens and hyphenated products such as HTTP, API, react-markdown, version strings, and caller-provided identifiers - keeps caller-supplied .NET technical identifiers case-insensitive regardless of the caller's set comparer - mirrors the direction policy across TypeScript batch and streaming APIs, Kotlin/Android, Swift/Apple, Rust, and C#/.NET - adds two canonical fixtures, bringing the shared corpus to 932 cases, plus direct platform regression tests - documents the unavoidable acronym/product-name heuristic boundary ## Root cause The technical-token filter treated any hyphen and every all-capital word as an identifier signal. Because those exclusions removed only Latin LTR evidence, ordinary English prose could be silently biased toward RTL when mixed with Persian, Arabic, or another RTL script. The first uppercase fix also needed a tighter boundary so an acronym-only phrase was not mistaken for emphasized prose. ## Developer impact Pure LTR input remains a no-op. Existing known technical terms and short acronym-only phrases remain excluded from language-majority decisions. Mixed blocks now retain genuine prose evidence, so base direction and opposite-direction isolation are more accurate without changing alignment policy or mutating source text. ## Tradeoff Shape alone cannot perfectly distinguish every acronym, emphasized word, and product name. The rule is deterministic and conservative: uppercase prose requires at least one capitalized word longer than the acronym bound, and ambiguous domain-specific names remain configurable through technicalIdentifiers or explicit direction. ## Validation on current head 416022d - pnpm run check — 16 files / 400 tests; 92.00% statements, 85.92% branches, 94.77% functions, 94.84% lines; 932-case corpus, docs, all packages, and Action bundle passed - JavaScript focused regression — 113 tests passed; character-by-character stream probes matched batch for acronym-only, mixed-case acronym, uppercase-prose, and hyphenated-prose cases - Android :core:test passed with JDK 21 - Windows native verification passed 1,166 assertions across all 932 fixtures (run locally on .NET 10 with major runtime roll-forward because exact CI-pinned .NET SDK 8.0.423 is not installed here) - git diff --check passed - exact hosted CI passed all 18 CI jobs and all five CodeQL language jobs on this commit; the PR is ready for review ## Security and language review The new Persian fixtures are generated policy evidence and are explicitly not native-speaker-certified. Source text remains in logical order, pure-LTR behavior remains on the identity path, and no new bidi control interpretation is introduced.