Name union arms from their schema node instead of their position #1409

Taken out of #1392 at review so that PR only contains code the pipeline actually runs. @KonstantinMirin wrote it.

What it is

scripts/union_arm_names.py names a root-level union arm from its schema node instead of its position:

1. Classify first. An arm that adds no new property is a validation constraint and creates no class. Bare $ref arms and scalar arms also create no class.
2. Name it using the first of these that is unique among its siblings:
   • the arm's title;
   • the parent name plus the PascalCased discriminator const;
   • the parent name plus the arm's full sorted required set.
3. Otherwise refuse with UnnameableArmError. It never falls back to a number.

tests/test_union_arm_names.py checks the rule against the pinned bundle. That includes the two response arms it can't name today.

Results on the pinned bundle: 557 root-level non-null arms.

• 169 named by title, 82 by const, 27 by required set.
• 278 create no class: 96 ref, 10 scalar, 172 validation.
• 1 refused: content-standards/get-content-standards-response.json oneOf[0], which has no title. Upstream fix: Both oneOf arms of get-content-standards-response carry no title, so generators name them by position adcp#7913.

Where the code is

On branch claude/issue-1409-union-arm-naming: #1392's head plus nothing else, before the extraction commit. It came from commits fcbf4229 and 5d705ca8. Those commits touch only scripts/union_arm_names.py and tests/test_union_arm_names.py. The design write-up is the "Union arm names" section of docs/design/type-generation.md on that branch.

What landing it needs

The rule isn't wired in yet. Arms are still named by position in restore_response_variant_aliases (f"{self.base}{index}"), using a hand-written 22-row table and a hardcoded BuildCreativeResponse reordering. Wiring it in is a one-line swap at class_names. The rest is two decisions about the public surface:

• The unnameable arm. If it's still unnameable, the build goes red as soon as the swap goes in. Either wait for Both oneOf arms of get-content-standards-response carry no title, so generators name them by position adcp#7913, or deliberately take that class off the flat surface and leave it reachable only by its domain path.
• The 38 publicly exported names ending in numbers become a breaking rename. That needs a deprecation shim and a migration table, ideally in a 9.x beta.

Also: test_no_derived_name_is_positional checks the rule against the JSON schemas, not the generated tree. A test that reads the generated classes has to land with the wiring.