test(depgraph): pin computeLevels on a diamond and on a value cycle

[thymikee]

Summary

Follow-up to the non-blocking points in the #3275 review.

• computeLevels relies on genPostorder yielding each node after its successors. No test checked this. A new test builds a diamond with a shortcut (a→b→d, a→c→d, a→d), plus a type-only back edge and an isolated file, and asserts the exact level map.
• On a value-import cycle, the edge that closes the loop now adds no height. Before refactor(depgraph): run graph traversals on @statelyai/graph #3275, height() returned 0 for the node in progress, so that edge added one. Levels in graph.json therefore differ from the old output only while an R4 cycle exists. A second test pins the current behaviour.
• The Tarjan pass in scripts/layering is already listed in Architecture quality: replace custom graph/layering code with maintained tools, close guardrail gaps, fix collocation (umbrella) #3276.

Validation

• pnpm check:affected --run passes.
• Iterating valueGraph.nodes instead of genPostorder makes both new tests fail.
• The /cycle/ match in the existing cycle test fits the library's error, getTransitiveReduction: the graph contains a cycle — ….

🤖 Generated with Claude Code

[cubic-dev-ai]

No issues found across 1 file

_{View guided diff | Turn on auto-fix | Re-trigger cubic}

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.12 MB	5.12 MB	0 B
Package (unpacked)	5.11 MB	5.11 MB	0 B
Package (download)	1.54 MB	1.54 MB	+3 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	24.1 ms	23.8 ms	-0.3 ms
CLI --help	74.4 ms	72.3 ms	-2.1 ms

[thymikee]

This PR is ready to merge: the new computeLevels tests in scripts/depgraph/model.test.ts look correct at 0f3d17a.

I did not run the tests locally. I checked the expected diamond and cycle values by reading the code and tracing them by hand.

Not blocking: the cycle assertion (a=1, b=0) at line 306 follows the sorted node order, so starting from b would give b=1, a=0. You could assert only that every level is finite and that z is greater than a, or leave it as is.

Smoke Tests are still running. The diff only touches a depgraph unit test, which the device smoke tests do not exercise. Wait for CI to finish before merging. There are no conflicts.

[thymikee]

Took the cycle-order point in b76abe0, but kept the assertion exact. The two cycle members must take levels {0, 1} in either order, and z must be 2. Checking only "finite and z > a" would also pass under the old behaviour, where the closing edge added one level ({1, 2} and z = 3). Making the closing edge add a level again fails this test. pnpm check:affected --run passes.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-07 18:49 UTC