[Miodec]

Roadmap step EP3. Redoes #7787 on top of the render store, reusing its RTL character table.

Summary

Each word resolves its direction once, from its first strong character. Rendering and the caret read that value instead of checking per letter.

 TestWords.push(word)
+  word.direction = first strong character, else the test's direction

 <Word>
-  classList wordRtl (only when the test marks RTL words)
+  dir={word.direction}, unicode-bidi: bidi-override

 caret.goTo
-  isWordRightToLeft(word, per letter)
+  testWord.direction ?? zen word's last typed character ?? test direction
   tape offset: still from the test direction

Numbers, newlines, tabs and Korean status stay test-level, but are scanned from the generated words plus the wordset, because words stream in during a test. koreanStatus stays one boolean in the EventLog context, so no version bump.

Evidence

Worst caret distance from the next letter's nearest edge, in px, same script on both commits:

Word that differs from the test	Tape	Before (0e07023d4)	After
Hebrew/Arabic in an English test	letter	723–736	79–89
Hebrew/Arabic in an English test	word	621–634	85–89
English in a Hebrew test	letter	714	105
English in a Hebrew test	word	606	105
Either	off	2–3	2–3

With tape off, __mt.caretOffset() on those words goes from 26–30px (caret on the wrong side of the letter) to 2–3px.

Unit tests: new words-generator.spec.ts covers facts found only in the wordset. test-words, strings, Words and test-ui-words specs are updated for per-word direction. Lint and type check pass.

Browser checks on the branch: Arabic and Hebrew in all tape modes, polyglot, backwards, joining scripts, facts that only appear late in the wordset, streaming, no visible flip on restart, and an 18ms scan on a 450k word list.

Known gaps:

• With tape on, the caret is still off its letter in the mixed-direction cases above. Up to ~194px was seen with other text. This is already there before this PR and is recorded under T3.
• Not yet compared with master: in zen, RTL input in an LTR test leaves the caret one letter off; with backwards, the caret is ~1,069px off after the first restart, until typing or the next restart.
• مرحبا123 shows its digits as 321.

Deviations from the plan: dir plus a Tailwind class rather than .ltr/.rtl; \p{N} rather than a numeral table.

Merge Danger

Door: two-way

Only rendering and caret code changes, with no stored data or EventLog version change.

Blast Radius: test-page

Direction of every word and caret placement on the test page, most visibly RTL languages, polyglot, custom text and the backwards funbox. .wordRtl is removed from test.scss, so custom user styles targeting it stop matching.

🤖 Generated with Claude Code

Co-authored-by: Nad Alaba 37968805+nadalaba@users.noreply.github.com

[Copilot]

Note

Copilot was unable to complete this review at the selected tier due to model policy or plan restrictions.

🟡 Changes recommended

A moderate active issue remains where getTestWordFacts needlessly allocates a giant joined string and its ASCII-only digit check can miss non-ASCII numerals, potentially leaving wordsHaveNumbers unset and the keymap top row hidden.

1 open issue

• Non-ASCII numerals are ignored in word fact detection · New

What changed in this PR

This PR implements EP3 changes to resolve direction per word (from the first strong character) and then uses that resolved value throughout rendering and caret logic, while also computing test-level “word facts” (numbers/newlines/tabs/Korean status) from generated words plus the full wordset to stay correct during streaming.

Changes:

• Resolve and persist direction on each test word; propagate through the render store and UI.
• Render words using per-word dir plus unicode-bidi handling; remove the .wordRtl styling pathway.
• Update caret/pace-caret to consume testDirection + per-word direction rather than per-letter RTL checks.
• Add/update unit tests and update docs/specs for the new per-word direction and word-facts behavior.

Key review notes (unresolved):

• unicode-bidi: bidi-override applied at the whole-word level can reverse digit sequences inside RTL words (e.g., "مرحبا123" displaying as "321"); consider unicode-bidi: isolate or conditional override use. (moderate)
• getTestWordFacts currently flattens the entire wordset into one string (extra allocation) and hasNumbers only checks ASCII \d, missing non-ASCII numerals; scan per-word with early exits and use /\p{N}/u. (moderate)

File	Notes
frontend/​static/​funbox/​backwards.css	Removes/adjusts bidi-related styling now handled per word.
frontend/​src/​ts/​utils/​strings.ts	Refactors direction utilities (e.g., getWordDirection / reversal).
frontend/​src/​ts/​utils/​direction-regex.ts	Introduces shared “strong RTL character” regex/table.
frontend/​src/​ts/​test/​words-generator.ts	Adds getTestWordFacts; moderate concern on full-wordset flattening and numeral detection.
frontend/​src/​ts/​test/​test-words.ts	Stores resolved per-word direction on test words.
frontend/​src/​ts/​test/​test-ui.ts	Updates restart/render integration to pass word objects (not just strings).
frontend/​src/​ts/​test/​test-logic.ts	Wires new direction/facts flow through test setup/streaming.
frontend/​src/​ts/​test/​pace-caret.ts	Updates caret API usage to rely on testDirection.
frontend/​src/​ts/​test/​caret.ts	Updates caret targeting to use per-word direction + testDirection.
frontend/​src/​ts/​states/​words-render.ts	Render-store buffering updated for word objects incl. optional direction.
frontend/​src/​ts/​states/​test.ts	Switches to explicit test direction helper (removes per-letter check path).
frontend/​src/​ts/​elements/​caret.ts	Caret positioning updated to use test/word direction inputs.
frontend/​src/​ts/​components/​pages/​test/​Word.tsx	Adds per-word dir + bidi handling; moderate concern re: digit reversal with bidi-override.
frontend/​src/​styles/​test.scss	Removes .wordRtl styling (behavior moved to per-word rendering).
frontend/​__tests__/​utils/​strings.spec.ts	Updates tests for new direction utilities/semantics.
frontend/​__tests__/​test/​words-generator.spec.ts	Adds coverage for word-facts computed from wordset + generated words.
frontend/​__tests__/​test/​test-words.spec.ts	Updates tests for per-word direction resolution behavior.
frontend/​__tests__/​test/​test-ui-words.spec.tsx	Updates UI word tests for object-based words/direction.
frontend/​__tests__/​components/​pages/​test/​Words.spec.tsx	Updates component tests for new word shape/render behavior.
frontend/​__tests__/​components/​pages/​test/​TypoHints.spec.tsx	Updates tests/helpers for new word object shape.
docs/​specs/​solid-monkeytype/​spec.md	Documentation updated to reflect per-word direction + word-facts scan.
docs/​specs/​solid-monkeytype/​roadmap.md	Roadmap updated to mark EP3 progress/completion notes.

_{Review effort: Balanced}