[polina-c]

This PR adds conformance cases for three behaviours that came up in the reviews of #3326 and #3327. No suite covers them today. The Dart core passes the new timing case since #3345. Each SDK's results are below, and the cases an SDK still fails on main are listed as known gaps in its harness.

The cases

1. core/validator_v0_8.yaml gets two v0.8 counterparts of test_v09_topology_plain_string_property_not_treated_as_child_ref. In both, a Text has usageHint: h1 and another component has the id h1.
   • test_v08_topology_plain_string_property_not_treated_as_child_ref expects the error for the orphan h1.
   • test_v08_topology_plain_string_property_does_not_close_cycle expects root -> h1 -> title to pass. If usageHint were read as a reference, it would close a cycle.
2. core/validator_v0_9.yaml and core/validator_v1_0.yaml each get two cases. Both use a custom Stack whose slots property is an array of ComponentId. The property name matches no naming convention.
   • test_v09_topology_component_id_array_dangling_reference_error and its v1.0 counterpart expect a dangling-reference error when the only item names no component.
   • test_v09_topology_component_id_array_child_reachable and its v1.0 counterpart expect the component an item names to be reachable from the root.
3. core/message_processor_scaling.yaml gets test_scaling_process_single_component_messages_composition_v10. It times n messages of one Text each into a root Column that names all n, under the relaxed config.
   • It uses a new fixture, conformance/test_data/catalogs/composition_catalog_v1_0.json. Its Column declares allowedParents and allowedChildren, and its Text declares allowedParents.
   • The sizes are 1000 and 4000. At 200 and 800, the extra cost in perf(web_core): check updateComponents incrementally under deferred graph checks #3327 gives a ratio of only 5.81, under the limit of 8.

Results

Every harness in the table reads the shared suites from conformance/, so the new cases run without harness changes. In Dart:

1. dart/a2ui_core/test/conformance/validator_conformance_test.dart runs validator_v0_9.yaml and validator_v1_0.yaml, so the new ComponentId array cases run in dart test.
2. dart/a2ui_core/test/conformance/message_processor_scaling_conformance_test.dart runs message_processor_scaling.yaml, so the timing case runs in dart test -P performance. The dart_a2ui_core (stable, performance) job of Flutter CI runs that.

Python is the only harness this PR changes, to list the cases it fails on main as known gaps.

Harness	v0.8 plain string	ComponentId array	Composition timing
Python core, main	Both fail, listed as known gaps; #3326 fixes them	Dangling cases fail, listed as known gaps, #3326 fixes them; reachable cases pass	Not run: the harness skips the suite until #3326
Python core, #3326	Pass	Pass	Passes
TypeScript web core, main	Pass	Pass	Not run: no timing runner until #3327
TypeScript web core, #3327	Pass	Pass	Fails: ratio 11.92; #3327 needs the change in its review
Dart core, main	Not run: a2ui_core does not implement v0.8	Pass	Passes since #3345 (ratio 14.9 before it)
Swift core	Not run: no v0.8 validator cases	Pass	Not run: no timing runner
Kotlin legacy agent	Pass	Pass (v0.9 only)	Not run: no timing runner

The timing numbers are median times on the same machine, at 1000 and 4000:

1. TypeScript web core with perf(web_core): check updateComponents incrementally under deferred graph checks #3327: 32.8 ms and 390.4 ms.
2. Dart core before perf(dart/core): check composition against what an update changes #3345: 302 ms and 4494 ms.

#3345 adds a Dart-only timing test for the same container, and says the case can move into the shared suite once #3326 and #3327 land. This PR adds it to the shared suite now. The Python core harness skips the whole suite on main, and the TypeScript harness has no timing runner there, so the case runs only in Dart until those PRs land. Each of them then needs to pass it or list it.

The #3345 shortcut applies only under a validation config that defers the graph checks. A processor with no config, or a strict one, still checks every reference on the surface for each message once a catalog declares constraints. The shared case uses the relaxed config, so it doesn't cover that path.

Known gaps

1. Python core: on main, the topology checks read references using the ids the surface holds.

   • A v0.8 string property counts as a reference when its value names a held component.
   • An array of ComponentId counts only when one of its items names a held component.

   perf(python/core): check updateComponents incrementally under deferred graph checks #3326 reads both from the catalog alone and passes all four cases, so it should drop the four SKIP_TEST_NAMES entries added here.

2. TypeScript web core: main skips the timing case, so nothing is listed. perf(web_core): check updateComponents incrementally under deferred graph checks #3327 adds a timing runner, which fails this case until it checks a parent's allowedChildren only against the components of the batch, as perf(python/core): check updateComponents incrementally under deferred graph checks #3326 does.

This PR is a draft because the v0.8 cases decide a question that #3326 raises: whether a property that is not a reference field becomes one when its value names a component. They follow the v0.9 and v1.0 cases and #3326: it does not. If main's reading is intended for v0.8, those two cases should go.

Testing

1. uv run pytest python/a2ui_core/tests/test_conformance.py conformance/tests, after merging main: 642 passed, 17 skipped.
2. In dart/a2ui_core, after merging main, dart test passes (1826 tests). dart test -P performance passes, the new case included, in three runs out of three.
3. dart test in dart/a2ui_agent passes (1105 tests).
4. node tests/conformance/conformance_test.mjs in typescript/core/web_core: passes.
5. swift test --filter ValidatorConformanceTests in swift/core: passes.
6. ./gradlew test --tests '*ConformanceTest*' in kotlin/agent_sdk_legacy: 173 tests pass.
7. The perf(python/core): check updateComponents incrementally under deferred graph checks #3326 and perf(web_core): check updateComponents incrementally under deferred graph checks #3327 results come from their branches, run against these suites.

_{🤖 Agent-generated}

[gemini-code-assist]

Code Review

This pull request introduces several new conformance tests across different protocol versions (v0.8, v0.9, and v1.0) to validate topology rules, specifically ensuring that plain string properties are not incorrectly treated as child references and that array properties containing ComponentId items are properly validated. Additionally, a new scaling performance test and its corresponding test catalog have been added to measure message processing under composition constraints. Finally, the Dart and Python test suites have been updated to skip these new tests where they are currently known gaps. As there are no review comments, I have no additional feedback to provide.