feat(types): Sequence[T] on response-only list fields for covariant adoption

[bokelley]

Refs #624

Adopters who subclass SDK response models and narrow the element type (e.g. list[MyPackage] instead of list[Package]) get mypy[assignment] errors because list[T] is invariant. This PR changes four response-only container fields to Sequence[T] — covariant in its element type — so Sequence[MyPackage] is a valid subtype of Sequence[Package] and the type: ignore is eliminated.

What changed

Four response-side fields (Category A from @bokelley's analysis in #624):

Field	Model	Type change
affected_packages	UpdateMediaBuyResponse1	list[Package] → Sequence[Package]
media_buys	GetMediaBuysResponse	list[MediaBuy] → Sequence[MediaBuy]
packages	MediaBuy (in GetMediaBuysResponse)	list[Package] → Sequence[Package]
media_buy_deliveries	GetMediaBuyDeliveryResponse	list[MediaBuyDelivery] → Sequence[MediaBuyDelivery]

Implementation path — post_generate_fixes.py (not hand-edited generated files):

• New RESPONSE_SEQUENCE_FIELDS config list + rewrite_response_list_to_sequence() function rewrites the annotation and inserts from collections.abc import Sequence on every codegen run.
• generate_ergonomic_coercion.py updated to detect Sequence[T] fields (via expanded is_list_of) and emit Sequence[T] (not list[T]) in the _patch_field_annotation call so _ergonomic.py's runtime coercion stays consistent with the source annotation.
• _ergonomic.py regenerated: media_buy_deliveries coercion patch now uses Sequence[MediaBuyDelivery] and from collections.abc import Sequence is imported.

Explicitly out of scope per @bokelley's analysis:

• Category B/C (required→optional field widening, account, idempotency_key, canceled): dissolve when Typed dispatcher rejects valid request when account is omitted on auth-resolved tools #623 lands — not an inheritance-typing problem.
• Category D (Literal narrowing, ~3 cases): possible Pydantic mypy plugin bug, separate investigation.
• geo_countries_exclude ×4: lives in core/targeting.py TargetingOverlay which is used in both request and response context — changing it could break adopters who .append() on request-side targeting lists.
• creatives (delivery): not yet located + same request/response ambiguity concern.
• Request-side list fields (packages/creatives on request types): explicitly excluded because adopters call .append() on those.

Bundled copies (src/adcp/types/generated_poc/bundled/): still use list[T] — these are not re-exported from _generated.py for these three response types and are not reachable from from adcp.types import .... If bundled types are ever promoted to the public API they would silently re-introduce the problem.

What was tested

• pytest tests/ -q (excluding test_real_tls_handshake_still_validates_hostname — pre-existing network failure that also fails on main): 4159 passed
• mypy src/adcp/: Success — no issues in 783 source files
• ruff check src/: All checks passed

Pre-PR review

• code-reviewer: approved — blocker (_ergonomic.py patching media_buy_deliveries back to list[T]) fixed via generate_ergonomic_coercion.py + _ergonomic.py regen; nits: bundled copies still use list[T] (explained above), \s+ vs \s* in regex (low-risk nit, field block is always multi-line in codegen output)
• dx-expert: approved — Refs #624 (not Closes) confirmed correct; _ergonomic.py comment updated to reflect which fields are migrated vs. still pending

Triage-managed PR. This bot does not currently iterate on
review comments or PR conversation threads (only on the source
issue). To unblock:

• Push fixup commits directly: gh pr checkout <num> →
  fix → push.
• Or re-trigger: comment /triage execute on the source
  issue.

See adcp#3121
for context.

Session: https://claude.ai/code/session_019aVaC1DXjVH6Kan4gBe1sb

---

Generated by Claude Code