fix(types): add UpdateMediaBuyResponse1 ergonomic coercion

[bokelley]

Summary

Replaces #857 on a repo-owned branch so required base-repo checks can run normally.

Follow-up to #854.

This extends the generated ergonomic coercion pass to include the update-media-buy success response arm. UpdateMediaBuySuccessResponse / UpdateMediaBuyResponse1 now gets the same package-subclass and media_buy_status string coercion treatment that CreateMediaBuyResponse1 already had.

What changed

• Added UpdateMediaBuyResponse1 to scripts/generate_ergonomic_coercion.py response analysis.
• Regenerated src/adcp/types/_ergonomic.py so affected_packages, packages, and media_buy_status are patched at import time.
• Added regression coverage for package subclass inputs on both update response package fields and string media_buy_status coercion.

Local validation

• PYTHONPATH=src pytest tests/test_type_coercion.py tests/test_code_generation.py -q --tb=short
• PYTHONPATH=src pytest tests/test_type_aliases.py tests/test_public_api.py tests/type_checks -q --tb=short
• PYTHONPATH=src ruff check scripts/generate_ergonomic_coercion.py src/adcp/types/_ergonomic.py tests/test_type_coercion.py

[aao-ipr-bot]

LGTM. Follow-ups noted below. Right call to mirror the CreateMediaBuyResponse1 coercion shape on the update-side success arm — the _find_media_buy_success_variant(module) refactor generalizes cleanly and _ergonomic.py:505-528 matches the upstream field declarations exactly (Sequence[Package] for affected_packages, list[Package] for packages, MediaBuyStatus for media_buy_status).

Things I checked

• Patch annotations mirror the generated source: update_media_buy_response.py:31-34 declares affected_packages: Sequence[Package] | None, packages: list[Package] | None, media_buy_status: MediaBuyStatus | None — _ergonomic.py:509-528 preserves the Sequence vs list distinction, which is load-bearing per the file preamble.
• _find_media_buy_success_variant cannot false-positive: only *Response1 in either module carries media_buy_id (verified across both create_media_buy_response.py and update_media_buy_response.py).
• Container detection in the comment generator (get_origin(base_ann) is AbcSequence) tracks post_generate_fixes.rewrite_response_list_to_sequence/widen_extension_point_lists_to_sequence outputs — the four list[X] → Sequence[X] comment churn sites at PackageRequest.creatives, CreateMediaBuyRequest.packages, ListCreativesResponse.creatives, GetMediaBuyDeliveryResponse.media_buy_deliveries are correctness, not cosmetic.
• Legacy status → media_buy_status bridging (media_buy_status_helpers.MEDIA_BUY_LEGACY_STATUS_VALUES) is preserved: the @model_validator(mode='before') at update_media_buy_response.py:37-57 runs ahead of the new field-level BeforeValidator, so the normalizer fires first.
• Conventional-commit prefix is right: fix(types): is non-breaking and additive. No public-API removal.
• Public surface unchanged: UpdateMediaBuyResponse1 / UpdateMediaBuySuccessResponse were already exported (__init__.py:613,616, aliases.py:214,607); this PR only loosens input validation, no JSON-schema export change.

Follow-ups (non-blocking — file as issues)

• UpdateMediaBuyResponse1.packages patches a codegen artifact, not a wire field. Per ad-tech-protocol-expert: the 3.1.0-beta.3, 3.0, and 2.5 update-media-buy-response schemas define only affected_packages on the success arm — packages at update_media_buy_response.py:32 has no upstream basis. The patch is harmless (extra='allow' already accepts it, no wire round-trip change) but advertises a field the spec doesn't define. Either fix the codegen to drop the spurious packages, or drop the patch + the corresponding test assertion. Cross-refs: adcontextprotocol/adcp#4906, #4895.
• Document the exclude=True discipline for subclass-list coercion. coerce_subclass_list widens input variance, but model_config = ConfigDict(extra='allow') on the response means subclass fields without Field(..., exclude=True) will leak into model_dump(). The new tests use it correctly; the docstring on coerce_subclass_list could note it. Same risk already exists on CreateMediaBuyResponse1.packages — not introduced here.

Minor nits (non-blocking)

1. Sequence coverage in the new test. tests/test_type_coercion.py:454-468 exercises affected_packages with a plain list, which already worked under the old list[Package] annotation. Passing a tuple (or any non-list Sequence) would actually prove the covariant-Sequence intent of the patch.
2. Hoist _find_media_buy_success_variant. It's a closure inside generate_code() (scripts/generate_ergonomic_coercion.py:267). Promoting it to module scope alongside get_base_type would make it independently testable. Cosmetic.

Approving on the strength of the protocol-shape check plus the Sequence/list parity with the generated source.