test(models): validate hot and offline catalogs independently

[HaningZS]

Summary

Addresses the catalog-test part of #652. CI suite execution already has an open contribution (#423); this PR deliberately does not duplicate it or close the whole issue.

The module contract explicitly describes the baked catalog as an offline fallback and the remote catalog as hot updates without shipping a release. Requiring them to be identical (including their provider keys) prevents that intended workflow. Instead:

• Both shipped files must survive the production parser without silent entry loss, sanitizing, truncation or deduplication.
• The shipped remote payload must update the actual picker overlay; clearing it must restore the offline lists.
• Existing model-id safety, malformed-payload, version-filtering and baked-model expectations remain in place. No model data or production code changes.

Before

Upstream f44b986a: the existing remote-catalog suite reports 17 pass / 1 fail. This is actual node:test output rendered as an image, not a GUI screenshot.

After

The corrected contract suite reports 19 pass / 0 fail. This is a test-only change: the suite now independently validates both files and adds the real shipped-payload overlay/offline round trip, rather than keeping the contradictory equality assertion.

Validation

macOS / Node 24.13.0:

• Catalog suites: 27 passed.
• Full npm run test:focused: 966 tests, 965 passed, 1 platform skip, 0 failed.
• npm run typecheck and npm run build: passed.
• Final integration with the independent worktree-origin (fix(worktree): persist isolation origins for agent restore #670) and reflection (fix(memory): respect providers and quota cooldowns in reflection #671) fixes: 1004 tests / 1003 passed / 1 platform skip / 0 failed, typecheck/build passed, no merge conflicts. The three real feature commits are combined in a temporary checkout; its full suite uses no test substitution.
• Negative-control experiment: injected a newline into the real remote file's first model label at the require boundary. The new shipped-remote test failed as intended (18 passed / 1 failed). This checks that relaxing equality did not remove the malformed-entry guard.

Self-review

No new dependencies, OS-specific implementation, network test dependency, production changes or permission changes. The original payload parser remains unchanged. Existing version-filter tests continue to exercise remote entries. The before/after use the same upstream payload data; the assertion contract intentionally differs and is described above.