test: Improve unit test coverage across packages

[Defilan]

Summary

Current unit test coverage is low across several packages. While E2E tests cover integration scenarios, improved unit test coverage would help catch regressions earlier and make the codebase more maintainable.

Current Coverage

From CI output:

Package	Coverage
internal/controller	33.9%
pkg/cli	10.4%
api/v1alpha1	0%
cmd/*	0%
pkg/agent	0%

Proposed Improvements

1. Establish baseline - Document current coverage and set realistic targets
2. Prioritize critical paths - Focus on controller reconciliation and CLI commands
3. Add coverage reporting - Consider Codecov or similar for PR coverage diffs
4. Update CONTRIBUTING.md - Add guidance on test expectations for new code

Areas to Focus

• Controller reconciliation logic (highest priority)
• CLI command parsing and validation
• Catalog loading and model lookup
• CRD validation and defaulting
• Error handling paths

Notes

For early-stage projects with heavy K8s integration, E2E tests often provide more value than unit test percentages. This issue is about improving coverage incrementally, not blocking PRs on arbitrary thresholds.

[bsnux]

I can start working on this. Please, let me know if that works :)

[repairman29]

Hey there! This looks like something I could help with. I'd be happy to take a look and see what we can do.

(Found via Echeo)

[Defilan]

Controller coverage improved from 34.5% → 76.4% (75 new test specs). PR incoming.

Covered:

• Reconcile flow: not-found, model-not-found, model-not-ready, Deployment+Service creation, Metal accelerator skip, PVC creation, nil replicas default
• Pure functions: sanitizeDNSName, buildModelInitCommand (all 4 branches), buildCachedStorageConfig, buildEmptyDirStorageConfig, constructService, constructEndpoint, determinePhase, resolvePriorityClassName, resolveEffectivePriority
• ensureModelCachePVC: defaults, custom size, ReadWriteMany, StorageClassName, idempotency, invalid size
• Model controller: cache hit, httptest download success/failure, local copy success/failure, formatBytes, getLocalPath, computeCacheKey, checkAcceleratorAvailability, copyLocalModel, downloadModel

[Defilan]

CLI coverage improved from 9.1% to 35.8% — added tests for all command constructors, pure functions (buildInferenceService, applyCatalogDefaults, sanitizeServiceName, isLocalSourcePath, validateLocalPath, computeCacheKey, formatAge, formatDuration), resource routing logic, benchmark output formatters/parsers (parseNvidiaSMI, parseSweepValues, outputTable, outputMarkdown, outputStressTable/JSON/Markdown, outputComparison*, outputSweepTable, ReportWriter), catalog display functions (runCatalogList, runCatalogInfo), and version check I/O (readCache, writeCache, getCacheFilePath, compareAndNotify).

[Defilan]

Hey @bsnux and @repairman29, appreciate you both offering to help on this! I should have responded sooner, sorry about that.

I ended up tackling the controller and CLI coverage myself back in February. Controller went from 34% to 78% and CLI from 9% to 36%, so the biggest gaps there are mostly filled in.

There's still plenty of uncovered ground though if either of you are interested. pkg/agent is at 57% (the watcher and executor logic especially), cmd/metal-agent is at 6%, and the API types are at 0%. Any of those would be great to pick up. The agent package is probably the most interesting one since it manages llama-server processes on macOS.

Happy to point you at specific files or answer questions if you want to dig in.