test: pin Wan CausalConv3d streaming cache (PRPUNDIT-24)

[jiagaoxiang]

This PR was created by dougljia via Test Gap Resolver.

Summary

• One parameterized CPU suite covers both Wan 2.1 vae2_1.py::CausalConv3d.forward and Wan 2.2 vae2_2.py::CausalConv3d.forward.
• Chunked streaming reuses the last CACHE_T frames of the previous chunk as cache_x and must match a single full-sequence pass; an independent F.pad + Conv3d oracle pins the cache-splice / reduced-front-padding arithmetic.
• Follows linxwang's note that PRPUNDIT-24 consolidates both implementations into one parameterized streaming-cache test (do not file a second PR for vae2_2.py).

Closes test gap PRPUNDIT-24.

Test plan

• PYTHONPATH=. pytest tests/unit_tests/backends/diffusion/test_wan_vae_causal_conv3d.py (12 passed, CPU: 6 tests × 2 VAE modules)
• black --check on the new file

[Copilot]

🟢 Approval recommended

The change is isolated to a new test file and accurately exercises the existing CausalConv3d cache/padding behavior in both targeted modules.

Pull request overview

Adds a parameterized unit-test suite to pin CausalConv3d.forward streaming-cache behavior for both Wan 2.1 and Wan 2.2 VAE implementations, ensuring chunked streaming output matches a full-sequence pass and validating the cache-splice/front-padding arithmetic via an independent F.pad + F.conv3d oracle.

Changes:

• Introduces a single parameterized test module covering both vae2_1.py and vae2_2.py CausalConv3d implementations.
• Validates full-sequence output against an independent padding+conv oracle and asserts expected causal padding layout.
• Verifies chunked streaming with cached frames (CACHE_T) matches full-sequence output, including full/partial/no-temporal-padding cache cases.

File summaries

File	Description
tests/unit_tests/backends/diffusion/test_wan_vae_causal_conv3d.py	New parameterized unit tests that pin causal temporal padding and streaming-cache equivalence across Wan 2.1 and 2.2 VAE modules.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[jiagaoxiang]

TestIntel PR Steward: Checked in on this PR after the poll flagged CI failure + a new review.

CI — build-docker (3.12) failure: not caused by this change. The Docker image itself built fine (all 14 layers, mostly cached, naming to docker.io/tasimage/primus:pr-1086 done). The job then fails at the registry-push step:

docker login -u tasimage -p
flag needs an argument: 'p' in -p
##[error]Process completed with exit code 125.

The -p password argument is empty, which reads as the Docker Hub credentials secret not being resolved for this run — a known GitHub Actions restriction on fork-originated pull requests (secrets aren't passed to pull_request-triggered workflows from forks by default). This PR only touches tests/unit_tests/backends/diffusion/test_wan_vae_causal_conv3d.py, so there's no code path here that affects docker build/push. Since the root cause is a missing/unavailable secret rather than a transient runner issue, rerunning is unlikely to change the outcome — this needs a maintainer to either approve the workflow run with secrets or adjust the docker-build job to skip the push step for fork PRs. Leaving CI as-is rather than reworking unrelated docker-build/CI config from this PR.

Copilot review: recommended approval, 0 inline comments — nothing actionable to address.

Still waiting on a human approval before this can merge; will keep watching.