test(models): pin qwen3.8-max resolution per provider, not across all four - #11590

Part of the "every gate is red on every branch" cleanup — one of six independent unit-test failures on release/v3.8.51.

The red test

✖ chatCore lifecycle resolution rewrites the model before dispatch
  AssertionError: resolvedModel for qwen-cloud-token-plan
  actual: 'qwen3.8-max', expected: 'qwen3.8-max-preview'

The test's premise expired halfway

Its docblock states the model "ships everywhere as qwen3.8-max-preview
(bailian-coding-plan, qoder, qwen-cloud-token-plan, qwen-web)". Probed against the live
catalogs and hasKnownProviderModel:

Two of the four now list the bare id and no longer carry -preview at all.

The code is right

resolveModelAlias short-circuits the built-in alias when the provider serves the id
itself:

// The provider serves this id itself — nothing is deprecated from its point of view.
if (provider && hasKnownProviderModel(provider, modelId)) return modelId;

Rewriting qwen3.8-max → -preview on those two would dispatch an id they no longer
serve — the mirror image of the 404 the original fix prevented. That short-circuit is
also why the BUILT_IN_ALIASES entry correctly stays rather than being dropped: the
other two still need it, and the alias map's own comment anticipated exactly this split
("Drop this line if Alibaba ever ships a distinct GA qwen3.8-max").

What changed

The loop asserted one outcome for all four providers, which the split made impossible to
express. It now asserts both directions, and adds a case pinning the catalog premise
itself:

const SERVES_BARE    = ["qwen-cloud-token-plan", "qwen-web"];
const SERVES_PREVIEW = ["qoder", "bailian-coding-plan"];

So if a catalog flips again, the premise case fails first and names the provider that
moved, instead of leaving the lifecycle assertions failing with no explanation. The
docblock records the half-expiry rather than pretending the original premise still holds.

Verification

# base branch
✖ chatCore lifecycle resolution rewrites the model before dispatch
ℹ tests 4  ℹ pass 3  ℹ fail 1

# with this change
✔ bare qwen3.8-max resolves to the canonical -preview id
✔ the canonical id is a no-op through the alias map (no double rewrite)
✔ the alias target carries the real 1M window, not the 128k fallback
✔ the catalog split the two ids across providers
✔ chatCore lifecycle rewrites the model only where the provider needs it
ℹ tests 5  ℹ pass 5  ℹ fail 0

Test-only; no production code touched.