flake: custom_nodes_e2e (core:1/2, 2/5) fail on ComfyUI-VideoHelperSuite S3 tier across unrelated PRs

## Summary

`custom_nodes_e2e (core:1/2, 0)` and `custom_nodes_e2e (2/5, 0)` are failing on a shared signature across multiple unrelated open PRs, not reproducing on `main`.

## Evidence

- FE [#16311](https://github.com/Comfy-Org/ComfyUI_frontend/pull/16311) (`test(assets): cover view mode remount persistence`, touches zero custom-node/backend files) — run [33582207019](https://github.com/Comfy-Org/ComfyUI_frontend/actions/runs/33582207019) fails both shards before AND after a full rerun, identical signature both times.
- FE [#16432](https://github.com/Comfy-Org/ComfyUI_frontend/pull/16432) (unrelated diff, also touches no custom-node/backend files) — run [33588956774](https://github.com/Comfy-Org/ComfyUI_frontend/actions/runs/33588956774) fails the exact same two shards, same timeframe.
- Failure signature (identical on all four job instances checked):
```
[tier-pack] tier=S3 pack=ComfyUI-VideoHelperSuite result=fail
✘ browser_tests/tests/customNodes/allNodes.spec.ts:61:5 › all nodes by tier @custom-nodes › S3: enrolled registered-node save/reload outcomes match exact contracts
```
- `origin/main`'s last 5 "CI: Custom Nodes Ecosystem Matrix" runs (2026-09-02T03:16Z–06:06Z) are all green — this does not reproduce on main, only on PR-triggered custom-nodes-matrix runs.

## Impact

Not merge-blocking (`core:1/2`/`2/5` aren't required checks — `mergeable_state` stays `unstable`, GitHub merge queue still accepts affected PRs), but it's cross-PR noise that will cause reviewers/bots to burn reruns treating it as a per-PR regression.

## Suspected cause (unconfirmed)

Likely upstream ComfyUI-VideoHelperSuite pin drift or a flaky external fetch during the S3 tier's custom-node pack setup, not a frontend code issue — needs someone to bisect the pack/tier harness itself.

Start at browser_tests/tests/customNodes/allNodes.spec.ts:61 and inspect the custom-nodes-matrix setup for the S3 ComfyUI-VideoHelperSuite pack, comparing PR-triggered runs with the green origin/main runs. Reproduce or bisect the shared failure and trace the tier-pack setup; done means the unrelated PR shards no longer fail, or the external cause and required follow-up are clearly documented.

Written by the indexing model from the issue text.