Add failing tests: placeholder gate rejects valid reordered human translations

TranslationValidator compares a source's positional and sequential specifier
views independently, so it only permits reordering when the source is ALREADY
positional. A non-positional source (`%@ … %@`) whose translation reorders via
positional specifiers (`%2$@ … %1$@`) is rejected — even though that is the
standard, Apple-documented iOS reordering mechanism and is correct at runtime
(`String(format:)` honors positional specifiers regardless of the source shape).

The human-translation gate added for the String Catalog fold
(`CatalogStrings.trusted_human`, `PluralStrings.human_forms_for`) inherits this,
so it rejects and downgrades valid, currently-shipping human translations to
machine/English. This is not hypothetical: the committed translated `.strings`
contain 23 key-as-source strings across 34 locales (513 cells) that positionalize
a bare-`%@` English source this way, e.g. ar "%@ of %@ used on your site" =>
"%1$@ من %2$@ على موقعك".

Two failing tests pin the correct behavior:
- translation_validator_test.rb — the root cause at the validator.
- catalog_strings_helper_test.rb — the user-facing downgrade in the fold.

The fix belongs in TranslationValidator (shared by AI, plurals, and regular
human strings): treat a non-positional source as implicitly numbered 1..N by
appearance order and accept a fully-positional candidate whose index→type map
matches. Existing tests already keep genuine breakage (sequential flip, type/
count change) rejected, so a correct fix cannot just drop the distinction.

Author: oguzkocer

fastlane/lanes/catalog_strings_helper_test.rb

@@ -73,6 +73,21 @@ def test_human_translation_is_used_and_marked_translated
     assert_equal({ 'state' => 'translated', 'value' => 'Enregistrer' }, cell(cat, 'a', 'fr'))
   end
 
+  # A human translation that reorders a non-positional source (%@ … %@) via positional specifiers (%2$@ … %1$@)
+  # is valid at runtime — String(format:) honors positional specifiers regardless of the source's shape — and is
+  # the standard iOS reordering pattern for RTL / word-order-differing locales. It must ship as `translated`, NOT
+  # be rejected by the placeholder gate and downgraded to machine/English. This is the user-facing half of the
+  # TranslationValidator regression (see translation_validator_test.rb): 23 key-as-source strings across 34
+  # locales (513 currently-shipping cells) do exactly this, e.g. ar "%@ of %@ used on your site" folds to
+  # "%1$@ من %2$@ على موقعك". With no AI tier here, the rejected human currently degrades to English (needs_review).
+  def test_reordered_human_translation_of_a_non_positional_source_is_kept_as_translated
+    cat = catalog('a' => entry('%@ of %@ used'))
+    capture_stderr { fold(cat, translations: { 'fr' => { 'a' => '%2$@ sur %1$@ utilisés' } }) }
+
+    assert_equal({ 'state' => 'translated', 'value' => '%2$@ sur %1$@ utilisés' }, cell(cat, 'a', 'fr'),
+                 'a reordered human translation must ship as translated, not be downgraded')
+  end
+
   # A whitespace-only GlotPress value is not a real translation: it must not ship as `translated` — the key
   # should still reach the model and land as `needs_review`, same as if GlotPress had no value at all.
   def test_whitespace_only_human_value_is_treated_as_untranslated

fastlane/lanes/translation_validator_test.rb

@@ -23,6 +23,24 @@ def test_positional_type_change_is_rejected
     refute V.placeholders_match?('%1$@ posts', '%1$d posts')
   end
 
+  # A NON-positional source (%@ … %@) whose translation reorders via positional specifiers (%2$@ … %1$@) is the
+  # standard, Apple-documented iOS way to reorder arguments for target grammar: String(format:) honors positional
+  # specifiers regardless of the source's shape, and each specifier still reads an argument of the same type. So
+  # this must be ACCEPTED, exactly like the already-positional reorder above. The gate currently rejects it,
+  # because it compares the source's positional/sequential views independently and a bare-%@ source has an empty
+  # positional map. This is not academic: real GlotPress data has 23 key-as-source strings across 34 locales
+  # (513 shipping cells) that positionalize a bare-%@ English source this way — e.g. the Arabic
+  # "%@ of %@ used on your site" => "%1$@ من %2$@ على موقعك". The human-translation gate this catalog work adds
+  # (CatalogStrings.trusted_human, PluralStrings.human_forms_for) rejects every one of them and downgrades a
+  # valid, currently-shipping human translation to machine/English.
+  #
+  # (Genuine breakage stays rejected by the existing tests: a sequential flip `%@: %d` => `%d : %@` and a type
+  # change are still caught, so a correct fix cannot simply ignore the positional/sequential distinction.)
+  def test_positionalizing_a_non_positional_source_to_reorder_is_allowed
+    assert V.placeholders_match?('%@ - %@', '%2$@ - %1$@')            # two objects, reordered
+    assert V.placeholders_match?('%@. %d posts.', '%1$@. %2$d posts.') # object + int, positionalized in place
+  end
+
   def test_sequential_order_must_be_preserved
     refute V.placeholders_match?('%@: %d', '%d : %@') # flipped non-positional args
     assert V.placeholders_match?('%@: %d', 'Total %@: %d') # same order, prose changed