[pedrol2b]

Related to #22 (intentionally not using a closing keyword — see Notes).

Summary

Text recognition failed to build on iOS as soon as any single language model
was disabled through the selective ML Kit install, which is the entire point
of that feature. Reported against 2.0.0 with an Expo prebuild config, but the
bare Podfile path produces identical flags and fails the same way.

Root cause

DomainTextRecognitionLanguage declares all five cases unconditionally, while
every case in TextRecognitionServiceFactory.createTextRecognizer was
individually wrapped in its own #if MLKIT_TEXT_RECOGNITION* flag with no
default:. Disabling a language compiled its case out while the enum still
declared it, so the switch was no longer exhaustive:

TextRecognitionServiceFactory.swift:31:7: error: switch must be exhaustive

Only all-five-enabled compiled, or all-five-disabled (where
MLKIT_TEXT_RECOGNITION_ANY goes unset and the whole file drops out). Every
partial selection — the feature's actual purpose — was broken.

This shipped because no iOS CI job compiles a partial config: build-ios.yml
builds the example app, and example/ios/Podfile pins all five to true.

The fix

The #if guards move inside each case body, via a makeOptions(for:)
helper that returns nil for models excluded from the build. The switch lists
all five cases unconditionally, so it stays exhaustive under every flag
combination.

A language whose model was excluded now throws a descriptive error naming the
configuration key to enable, matching the IllegalStateException Android
already raises from its own factory. createTextRecognizer/create become
throwing, so HybridTextRecognizer.init throws and its single call site in
HybridVisionCameraMLKit uses try.

Alternatives considered and rejected

default: + fatalError (suggested in the issue) compiles, and Swift
emits no unreachable-default warning when all five are enabled. Rejected
because it crashes the process where Android throws a catchable exception for
the same misconfiguration, and because it permanently disables exhaustiveness
checking — verified by simulating a sixth enum case: the default: version
compiles silently, this version fails the build.

Gating the enum cases (the issue's alternative) does not work.
TextRecognitionLanguage+toDomain.swift maps the Nitro-generated
TextRecognitionLanguage — C++-backed, all five cases, marked DO NOT MODIFY —
so removing domain cases relocates the same error into the mapper. It also
would not surface at the JS/TS boundary, since that union is generated from the
Nitro spec and is independent of build flags.

Validation

Real pod install + xcodebuild per config, reading the emitted
SWIFT_ACTIVE_COMPILATION_CONDITIONS back from the generated xcconfig rather
than trusting the Podfile (the podspec merges over all-true defaults and does
not validate keys, so a typo would silently produce an all-enabled build).

Config	Branch	Scope	Result
latin only	develop	pod scheme	FAIL — switch must be exhaustive at 31:7
latin only	this branch	pod scheme	PASS
chinese only, latin off	this branch	pod scheme	PASS
all five	this branch	pod scheme	PASS
latin only	this branch	full app	PASS (links)

Rows 1 and 2 use an identical Podfile and identical emitted flags; the branch
is the only variable. Row 1 reproduces the reporter's error at the same file,
line, and column. Row 3 also confirms the podspec still pulls base
GoogleMLKit/TextRecognition for shared types when textRecognition => false.

Not covered: the runtime throw reaching JS. Low risk rather than unknown —
HybridVisionCameraMLKit.createTextRecognizer already threw a RuntimeError
for the disabled-feature case before this change, so that propagation path is
pre-existing and exercised.

Notes

• The issue must stay open. No closing keyword here, so merging to
  develop changes nothing. Note the fix commit body does contain one, which
  will auto-close iOS build fails with "switch must be exhaustive" when any text recognition language model is disabled #22 when it reaches main — to be amended before release.
• CI left unchanged by maintainer decision; the five ML Kit scripts are the
  complete set, so the enum will not grow.
• CHANGELOG entry added under [Unreleased]; version header stamped at release.