[bokelley]

Closes #1616.

Validating a schema-valid McpWebhookPayload or AsyncResponseData document raised a raw TypeError from inside pydantic instead of returning a model or a ValidationError:

TypeError: Unable to apply constraint 'min_length' to supplied value
  (pydantic/_internal/_validators.py min_length_validator: object of type 'Losses' has no len())

Cause

Not the allOf flattening, and not fix_legacy_purchase_accepted_losses (which only edits legacy_purchase_continuation_input.py). An array schema that carries an allOf of contains / if-then constraints has no Python shape codegen can give those keywords, so datamodel-code-generator renders the composition as an empty object model and widens the field to list[Item] | Empty. Pydantic then applies the array's min_length / max_length to the whole union, and a dict reaching the model arm trips the constraint with a TypeError rather than a ValidationError. Three generated fields had the defect:

• PurchaseContinuation1.losses: list[Loss] | Losses (media_buy/request_proposals_response.py, min_length=2)
• RefineProposalsRequest.refinements: list[ProposalRefinement] | Refinements (min_length=1, max_length=25)
• RefineProposalsResponse1/2.results: list[Results8] | Results (min_length=1)

Every schema-valid webhook or async-response document whose result union visited the RequestProposalsResponse arm while being resolved crashed the same way, which is why all four bundle examples of core/mcp-webhook-payload.json and the synthesized documents were refused.

Fix (generator rule, not the generated file)

scripts/post_generate_fixes.py gains drop_empty_model_arms_from_array_fields, a generic AST pass: for every annotation union that carries a list[...] arm together with an arm naming an empty AdCPBaseModel subclass (class X(AdCPBaseModel): pass), drop the empty arm, in either order, and leave the class in place. The class is a public adcp.types name (Losses, Refinements, Results are in _generated.py and the public API snapshot), so removing it would be a 9.x surface change for nothing; the public API snapshot is unchanged. The rule is idempotent and registered in post_generation_manifest.json as firing. It fired on exactly the three modules above; make validate-generated passes.

A dict on any of those fields is now a ValidationError (list_type), the list constraints still apply (losses with one entry is still refused by min_length=2), and the three empty classes keep validating as before for anyone who imported them.

Tests

• tests/test_constrained_array_empty_arms.py: all four bundle examples of core/mcp-webhook-payload.json validate, both as written (data) and in the completed form the sweep grades (the documents that raised); the sweep's synthesized core/async-response-data.json document resolves through the union; a dict on losses / refinements / results is a ValidationError; the legacy continuation still accepts and length-checks its losses; and an AST scan asserts the generated tree has no empty-model arm beside a list anywhere. 9 of the 14 fail against the unfixed tree.
• tests/conformance/test_wire_round_trip.py: core/mcp-webhook-payload.json and core/async-response-data.json are removed from REFUSING, so the sweep enforces the fix (the "only shrinks" assertion fails if they come back). New test_no_schema_valid_document_raises_outside_validation forbids any refusal recorded under an exception class name, with no allowlist; _round_trip.py names those classes (RAISED_OUTSIDE_VALIDATION) so the sweep and the test share them.
• python -m tests.conformance._round_trip on this branch: 1,008 graded documents, 5 refusals (all declared ValidationError refusals; the 6 TypeError ones are gone).

Validation

uv sync, make lint, make typecheck-all, make validate-generated green; the round-trip, schema/model parity, public API, codegen invariant, hierarchy and code-generation suites pass. The full CI-equivalent suite runs locally after #1657's; CI is the gate.

Based on main; it touches three regenerated modules that #1657 does not, and will be rebased onto main once #1657 lands so CI grades the combined tree before merge.

🤖 Generated with Claude Code

[aao-secretariat]

Ladon verdict: Approve

Approve — clean fix for #1616 with no blocking or medium findings.

What I checked:

• drop_empty_model_arms_from_array_fields in scripts/post_generate_fixes.py is a sound AST pass: strips the empty AdCPBaseModel: pass arm from unions that also carry a list[...] arm, with correct guards against emptying the union (not kept), no-ops (len(kept)==len(arms)), and non-list unions (has_list). Reverse-ordered offset splicing avoids overlap; col_offset is byte-safe (ASCII union lines).
• The three regenerated domain modules (refine_proposals_request.py, refine_proposals_response.py, request_proposals_response.py) match the transformation exactly (list[Loss] | Losses → list[Loss]) and are paired with the generator rule — not hand-edits inside domains/**, so the "generated code is not source" gate is satisfied.
• Semver signal: fix(types): is correct. No schema-valid document ever produced the Empty arm (array fields with min_length), public classes stay defined, snapshot unchanged, nothing that deserialized before breaks — not a breaking public-surface change.
• Tests are the right shape: tree-wide AST invariant, bundle examples, ValidationError-not-TypeError, REFUSING shrinks, new no-allowlist raise guard.
• Bulk of the diff is black paren-wrapping (semantically inert).

No Critical/High/Medium findings. gated_paths false, no high-risk paths, no no-auto-approve team match. Falls through to row 9 → approve. (Note: local CI gates ruff/mypy/pytest remain the floor and are presumed enforced in CI.)

[aao-secretariat]

Ladon verdict: Approve

Approve — clean pass, no blocking or medium findings.

Checked: the new drop_empty_model_arms_from_array_fields generator rule in scripts/post_generate_fixes.py correctly strips the codegen-artifact empty-model arm from list[...] | Empty unions (#1616) — it only drops empty AdCPBaseModel subclasses beside a list[...] arm, leaves public classes (Losses/Refinements/Results) defined and exported, is idempotent, and uses AST-surgical reverse-sorted bottom-up offset edits that keep replacement arithmetic safe. The three domain files are regeneration output, not hand-edits; pre-fix behavior was a raw TypeError crash, so narrowing to list[...] breaks no deserialization contract — fix(types): is the correct semver signal. New tests enforce the invariant tree-wide and lock the round-trip sweep against regression.

Decision path: no critical/high (row 1 n/a); gated_paths false (row 2 n/a); high_risk false (rows 3/5 n/a); no medium findings (rows 4/8 n/a); prior decision was approve (row 6 n/a); no no-auto-approve team match (row 7 n/a). Falls through to row 9 → approve. Consistent with the prior approve on the preceding head.