test: add comprehensive integration tests across modctl

[aftersnow]

Summary

• Add shared mock OCI registry (test/helpers/mockregistry.go) with fault injection (latency, partial response, status override, Nth-request failure, per-path faults) for httptest-based integration testing
• Add ~35 integration tests across 8 dimensions: functional correctness, network errors, resource leaks, concurrency safety, stress, data integrity, graceful shutdown, and idempotency
• Regression tests for the 4 bugs found during analysis, all fixed on main since this PR was opened: ReadCloser leak (bug: Push leaks ReadCloser from PullBlob on both success and error paths #491), splitReader goroutine leak (bug: splitReader goroutine can block indefinitely if pipe reader is abandoned #492), disableProgress data race (bug: disableProgress global variable has data race (no atomic/mutex protection) #493), auth error retry (bug: retry.Do retries 401/403 auth errors instead of failing immediately #494)

New Files (13 files)

File	Tests	Build Tag
test/helpers/mockregistry.go + _test.go	9 self-tests	default
test/helpers/tracking.go	—	default
pkg/modelfile/modelfile_integration_test.go	2	default
pkg/backend/pull_integration_test.go	11	default
pkg/backend/push_integration_test.go	8	default
internal/pb/pb_integration_test.go	2	default
cmd/modelfile/generate_integration_test.go	3	default
cmd/cli_integration_test.go	3	default
pkg/backend/pull_slowtest_test.go	4	slowtest
pkg/backend/push_slowtest_test.go	1	slowtest
pkg/modelfile/modelfile_stress_test.go	2	stress
pkg/backend/pull_stress_test.go	2	stress

Notes from the review round

• The TestKnownBug_* reverse-assertion tests were flipped into TestIntegration_* regression tests once main (merged in c38fe8d) fixed bug: Push leaks ReadCloser from PullBlob on both success and error paths #491, bug: disableProgress global variable has data race (no atomic/mutex protection) #493 and bug: retry.Do retries 401/403 auth errors instead of failing immediately #494.
• MockRegistry reads PATCH and PUT upload bodies before taking its mutex, so network I/O does not serialize unrelated requests.
• The generated plan and design documents were removed.
• The slowtest and stress suites are not run by CI. They were broken in the first version of this PR (a global fault hit the manifest fetch, which is outside the retry loop, and the goroutine-leak test counted idle keep-alive connections). Both suites pass now; see the test plan.
• Finding for a follow-up: every Pull/Push builds its own http.Transport in pkg/backend/remote and never closes its idle keep-alive connections. The CLI exits after one operation, so this only matters for library users that call the backend repeatedly. The stress test disables keep-alives on the mock server to measure real goroutine leaks only.
• Finding for a follow-up: ProgressBar.Add on an existing name aborts and drops the old bar. Under -race, 10 goroutines replacing 10 bars in a tight loop deadlock mpb while it syncs decorator widths. The concurrent-access test uses unique bar names to avoid that path; production code replaces bars rarely.

Test Plan

• go test ./... -count=1 — all default tests pass
• go test -race ./internal/pb/ ./pkg/backend/ ./test/helpers/ — clean; go test -race -count=20 ./internal/pb/ — clean
• go test ./pkg/backend/ -tags slowtest -timeout 900s — passes (about 150 s)
• go test ./pkg/modelfile/ ./pkg/backend/ -tags stress -timeout 600s — passes
• golangci-lint run — 0 issues

[gemini-code-assist]

Code Review

This pull request introduces a comprehensive integration testing suite for modctl, covering functional correctness, network resilience, concurrency, and stress scenarios. It includes a new mock OCI registry helper and resource tracking utilities. Review feedback identifies a logic error in a 'KnownBug' test that prevents it from failing when fixed, suggests renaming a misleadingly titled test, recommends optimizing the mock registry's performance by reducing mutex contention during I/O, and advises removing a redundant implementation plan file that duplicates the source code.

[bergwolf]

Please do not submit your AI's plan.md ;)

[aftersnow]

Removed in 7dbe732 (both the plan and the design document under docs/superpowers). Sorry for the noise.

[aftersnow]

@bergwolf @imeoer Thanks for the review. Summary of the update:

• Merged current main (c38fe8d). main has fixed the 3 bugs the reverse-assertion tests documented (bug: Push leaks ReadCloser from PullBlob on both success and error paths #491, bug: disableProgress global variable has data race (no atomic/mutex protection) #493, bug: retry.Do retries 401/403 auth errors instead of failing immediately #494), so those TestKnownBug_* tests now fail as designed. They are flipped into TestIntegration_* regression tests (7dbe732), and the error-path test asserts AssertClosed directly instead of the conditional Gemini flagged.
• Review items: renamed TestIntegration_Pull_TruncatedBlob to TestIntegration_Pull_CorruptedContentSameLength; MockRegistry reads PATCH/PUT bodies before taking its mutex; the plan and design documents under docs/superpowers are removed.
• The slowtest and stress suites are not run by CI and were broken in the first version (4774414 fixes them, details in the commit message). Both pass now, and the default suite is race-clean. A pb test that deadlocked mpb under -race is fixed as well.
• Two findings for follow-ups, not changed here: each Pull/Push builds its own http.Transport and never closes idle keep-alive connections (only matters for library users), and ProgressBar.Add replacing a live bar in a tight loop can deadlock mpb's width sync.
• Local checks: go test ./..., go test -race ./internal/pb/ ./pkg/backend/ ./test/helpers/, -tags slowtest, -tags stress, and golangci-lint all pass.

PTAL.