test(dashboard-api): cover unhealthy/warnings/retry/catalog regression gaps

[yasinBursali]

What

Adds pytest coverage for four dashboard-api / host-agent behaviors that were previously untested:

• _compute_extension_status's stale-window → "unhealthy" transition for user extensions.
• enable_extension's multi-service mixed-outcome warnings list (cross-svc accumulator).
• Host-agent enable-retry edge cases (no-hook, corrupted progress JSON, mid-install setup_hook status).
• Host-agent _handle_model_download regression shield for the missing/corrupt catalog → 500 contract.

Why

• The L194 unhealthy return for user extensions lacked a regression test for the stale-started path; only the no-progress-file path was covered.
• enable_extension accumulates per-svc warnings across multi-service enables, but only single-svc cases were tested. A regression that broke cross-svc warning attribution (warning gets attributed to the wrong svc, or main's failure spuriously suppresses dep warnings) would have shipped silently.
• The host-agent enable-retry path has three edge cases that aren't pinned by the existing TestEnableRetry class shipping in another in-flight PR: extensions without a post_install hook, progress files that became corrupt mid-install, and extensions retrying while still in a setup_hook mid-install state.
• The model-download endpoint's missing-catalog and corrupt-catalog responses returned 403 ("Model not in library catalog") rather than 500 ("Model catalog unavailable"), conflating a server-side state issue with a permission denial. The 500 split is shipping in another PR; this lands a regression shield against a future revert.

How

tests/test_extensions.py (+185 lines, 3 new test methods):

• TestExtensionLifecycleStatus::test_stale_started_unhealthy — fixture uses now-305s for updated_at (must be >300s to fall through the started-recency window AND <3600s to avoid _read_progress suppressing the file as stale-error-only). Asserts the unhealthy path returns "unhealthy" for user extensions whose container is stale-started + reporting unhealthy.
• TestEnableExtensionHookReturnHandling::test_multi_svc_pre_start_failure_on_dep_does_not_warn — dep's pre_start fails, main's all pass; asserts warnings == [] (terminal pre_start failure isn't accumulated).
• TestEnableExtensionHookReturnHandling::test_multi_svc_post_start_failure_on_main_only_warns_main — main's post_start fails, dep all pass; asserts exactly one warning naming main, none for dep.

tests/test_host_agent.py (+285 lines, 2 new test classes):

• TestEnableRetryEdgeCases (named to avoid collision with the in-flight TestEnableRetry class) — three cases for the retry path that exercise no-hook, corrupted-JSON-progress, and mid-install setup_hook states.
• TestModelDownloadCatalogUnavailable — three cases asserting 500 for missing catalog, 500 for corrupt catalog, and 403 preserved for clean catalog with model not listed.

All tests use existing helper patterns (_setup_user_ext, _patch_mutation_config, in-process HTTPServer(("127.0.0.1", 0), _mod.AgentHandler) wire pattern, monkeypatch for module attribute changes) and run in tmp_path to avoid global state pollution.

Testing

On upstream/main (current state):

• 4 new tests pass: stale_started_unhealthy + 2× multi-svc warnings + TestModelDownloadCatalogUnavailable case (c) clean-catalog-not-listed.
• 5 new tests fail-as-expected (DRAFT-only):
  • 3× AttributeError on _read_progress_status (TestEnableRetryEdgeCases — the helper is added by another in-flight PR).
  • 2× 403 != 500 (TestModelDownloadCatalogUnavailable cases a and b — the 500 split lands in another in-flight PR).

After the prerequisite branches merge, the rebase will turn all 9 tests green.

Pre-existing 12 baseline failures in the full pytest suite (Python 3.14 / asyncio compat) verified unchanged via stash-pop on the PR-Y worktree — none introduced by this PR.

Review

Critique Guardian: APPROVED. Falsifiability verified across all 9 cases (assertions on response shape, content, and warnings-list contents — not just status codes).

Known Considerations

• This PR is DRAFT and must merge after the in-flight host-agent enable-retry branch and the host-agent runtime-hygiene branch land. The 5 fail-as-expected tests are the mechanical safeguard.
• Class-name collision avoided: TestEnableRetryEdgeCases (this PR) vs TestEnableRetry (other branch).
• A small coverage gap not pursued here: simultaneous dep pre_start + main post_start failures combined into one call. Edge-of-edge case; could be a future addition.

Platform Impact

• macOS: no impact — pytest runs identically; mocks shield against GPU_BACKEND-driven branching.
• Linux: no impact — same.
• Windows-WSL2: no impact — same.

[Lightheartdevs]

Merge-pass recheck: holding this as draft/red.

The regression coverage is valuable, but the PR body says 5 tests are fail-as-expected until prerequisite branches land, and the live api check is currently failing. Please rebase after the prerequisite host-agent/runtime work is merged, make the full API job pass against current main, and then undraft for review.