LGTM. Schema-cache refresh with zero wire-shape delta; alias layer survives the renumbering churn by construction.

Things I checked

• No field-shape changes on the wire. ad-tech-protocol-expert byte-compared three bundled wire schemas (bundled/media-buy/create-media-buy-request.json, bundled/media-buy/get-products-response.json, bundled/creative/sync-creatives-request.json) between schemas/cache/3.1.0-beta.4/ and 3.1.0-beta.5/. Identical modulo _bundled.generatedAt, including required, top-level properties, and $defs keys. The empty SCHEMA_DELTAS.md is correct, not a tooling failure.
• Alias-layer integrity. Every numbered class imported by aliases.py (AccountReference1/2, CreateMediaBuyResponse1/2/3, Deployment1/2, Destination1/2, AuthorizedAgents1..6, CatalogFieldBinding1, Field1) is present at matching positions in the regenerated _generated.py. The renumber churn (Signal81 → Signal, PricingOption341 → PricingOption211, etc.) lives in unaliased flat enum classes — true renumbering churn per the CLAUDE.md warning, no semantic delta.
• chore: semver signal is correct. ADCP_VERSION is the protocol pin, not the SDK semver. aliases.py, capabilities.py, _ergonomic.py, _forward_compat.py, __init__.py all unchanged. No public Python API surface change.
• Beta.4 cache preserved. 675 files added under schemas/cache/3.1.0-beta.5/, zero modified or removed under schemas/cache/3.1.0-beta.4/. Downstream tests pinned to beta.4 still resolve.
• Layering rule held. Generated churn is confined to src/adcp/types/generated_poc/** and _generated.py. No non-allowlisted module gained an import from generated_poc.

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

• add_rootmodel_getattr_proxy swapped AST for regex and lost a robustness property (scripts/post_generate_fixes.py:543-600). The insertion point is computed as next_class.start() or len(source). If datamodel-codegen ever emits a module-level statement between a RootModel union class and the next class (e.g., FooAlias = FooUnion), source[:offset].rstrip() leaves the alias line intact and visually attaches the 4-space-indented def __getattr__ to it — IndentationError at module scope. I checked the current beta.5 generated tree and no RootModel union is followed by module-level code, so this is latent, not active. But the AST predecessor was robust to that shape; the regex version is not. Cheapest fix: keep AST for this one function — single per-file parse is small next to the rglob cost.
• fix_deprecated_rootmodel_fields skips the first line of the annotation block (scripts/post_generate_fixes.py:1515-1521). _find_class_field_block returns start = annotation_start = cursor + len(" root: "), which is mid-line. The line-clear loop's line_start < start: continue always skips that first line. For multi-line RootModel root: Annotated[... annotations (what codegen currently emits), deprecated=True is never on that line. If codegen ever produces root: Field(deprecated=True, ...) single-line, Pydantic rejects it and this fix won't strip the deprecated=True. Worth tightening to a line-coverage check that doesn't require strict-less-than against a mid-line offset.
• add_deprecated_field_metadata substring matcher loosened from Annotated[…]Field( to Field( substring (scripts/post_generate_fixes.py:282-285). Strictly more permissive than the predecessor. Currently fine — no deprecated field's annotation contains a Field(-suffixed identifier — but \bField\( would restore the prior precondition at near-zero cost.
• SCHEMA_DELTAS.md is overwrite-only by design (scripts/diff_generated_types.py:202), so the beta.3→beta.4 delta the previous version recorded is gone from git history's working tree representation. For audit purposes, future bumps should fold the per-version delta into CHANGELOG.md so the cumulative trail survives the next regen.

Minor nits (non-blocking)

1. add_rootmodel_getattr_proxy anchors on :\n to find the class header end (scripts/post_generate_fixes.py:573). Works because datamodel-codegen always emits class Foo(...):\n. A header ending in a trailing comment would break it. An assert or comment recording the assumption would make the dependency explicit.
2. Conventional-commit title only narrates half of the PR. chore: bump AdCP schemas to beta.5 covers the version pin and the regen; the 333-line refactor in scripts/post_generate_fixes.py is independent perf work and is mentioned in the body but not the title. Not a semver issue — both pieces are chore:/refactor: grade — but the next reader of git log --oneline won't see that this commit also touched the build script. Notable for an otherwise-clean schema-bump PR; future bumps benefit from splitting.

Safe to merge.