fix: report unknown script reading direction instead of guessing left-to-right

[andrew-polk]

What

isRTLScript now returns boolean | undefined, where undefined means "we genuinely do not know this script's reading direction". Consumers omit isRtl entirely rather than storing a fabricated false.

Why

Intl.Locale answers either ltr or rtl for every code it accepts — it has no way to say "no idea". So placeholder script codes were getting a confident left-to-right answer:

Code	Meaning	Languages in our data
Zyyy	undetermined script	358
Zxxx	unwritten (mostly sign languages)	165
Zzzz	uncoded script	28
Qaaa–Qabx	private use	via typed tag
unregistered (e.g. Xyzw)	not a real script	via typed tag

A sign language has no reading direction at all, so isRtl: false was an assertion we had not earned. With undefined, a consumer storing a writing system's direction can leave an existing setting (or the user's own choice) alone instead of overwriting it with a guess. Anyone wanting a hard boolean writes script.isRtl ?? false, which behaves exactly as before.

Also fixed

Script variant relationships are now derived from ISO 15924's own names (Aran is "Arabic (Nastaliq variant)", Syrj is "Syriac (Western variant)"), so a variant inherits its parent's direction:

• Western Neo-Aramaic (amw-Syrj) reported left-to-right despite Syriac being right-to-left. This is the one genuinely searchable bug fixed here.
• Sidetic (xsd-Sidt) is pinned right-to-left. CLDR marks it RTL, but ICU builds predating Unicode 16 report it left-to-right. The langtags API 1.4 data merged into this branch makes Sidetic searchable, so this now matters in the UI.

Derivation resolves: Aran→Arab, Syre/Syrj/Syrn→Syrc, Cyrs→Cyrl, Latf/Latg→Latn, Hans/Hant→Hani.

Behavior across the whole dataset

484 language-script pairs report RTL, 8075 LTR, 554 unknown.

Reviewer notes

• Verified against CLDR release-48-2 scriptMetadata.txt (RTL column, derived from Unicode Bidi_Class): 0 mismatches across all 179 script codes CLDR has an explicit verdict for.
• A regression test pins Todr (Todhri) as left-to-right — an earlier draft of this work wrongly forced it RTL, which would have misrendered Albanian.
• isRtl was already declared isRtl?: boolean, so undefined was always representable; this is not a new shape.

---

This change is 

---

Devin review

[andrew-polk]

Consulted Devin on 2026-07-27 21:39 UTC up to commit cc2261e20a5fbf972c8b83f6f8934b842eb52172.

It found no bugs. It raised one item to investigate and three informational observations.

• Investigate — whether the Sidetic right-to-left pin is dead code. Answered: it isn't. Every script code this change depends on is present in the pinned iso-15924@3.2.0 registry copy, and the passing test is itself the proof. Replied and resolved: #discussion_r3661087137
• Informational (not mirrored, all three assessed): that variant parent lookup is single-hop and derived from registry names (correct, and verified that no parent-of-a-parent chains exist); that isRtl being absent rather than false is a change to what external consumers see (correct — that is the intent of the change, and how to version it is going to the developer); and that deleting isRtl on the hook path is safe because deepStripDemarcation deep-copies first (correct, and independently confirmed during the local review).

CI (run-tests) passed. No other review bots are configured on this repo.

_{Claude Opus 5 (1M context)}

[andrew-polk]

Consulted Devin on 2026-07-27 21:55 UTC up to commit 9de4456582546870bca7935e7ddca8adbd2aa6fc (the branch was squashed and rebased onto main, so this is a fresh review of the rewritten history).

No bugs. One item to investigate and four informational observations.

• Investigate — whether the Sidetic right-to-left pin is dead code. Same finding as the previous review; already answered and resolved above, and that thread is still anchored correctly after the rebase since the file content is unchanged. Short version: all 12 script codes this change relies on are present in iso-15924@3.2.0, so the pin is reachable, and the passing test proves it.
• Informational (not mirrored, all four assessed): variant parent lookup is single-hop and derived from registry names (correct — verified no parent-of-a-parent chains exist); isRtl being absent rather than false changes what external consumers see (correct, and intended — flagged to the developer, and noted on BL-16593 for the Bloom side); the Sidetic pin and the Book Pahlavi mapping depend on the host ICU/Unicode version (correct — that is why the pin exists, and its comment says to delete it once ICU catches up); and deleting isRtl on the hook path is safe because deepStripDemarcation deep-copies first (correct, independently confirmed during the local review).

No other review bots are configured on this repo.

_{Claude Opus 5 (1M context)}