fix(catalog): exclude generic 128k default from combo context min() (#10734)

[Prudhvivuda]

Summary

• Combo context_length in /v1/models is the min of sourced member windows only (canonical window or a provider/model-specific token limit).
• getTokenLimit()'s generic 128000 catch-all (specific: false) is no longer treated as a known window, so a 500k combo is not advertised as 128k because an unsourced or generic-default member entered min().
• Explicit combo context_length overrides still win. Nested combo-refs inherit the same sourced-min rule.

Closes #10734

⚠️ base-red inherited: #9985

Test plan

• node --import tsx/esm --test tests/unit/combo-context-generic-default-10734.test.ts
• node --import tsx/esm --test tests/unit/combo-context-prefix-resolution.test.ts
• node --import tsx/esm --test tests/unit/models-catalog-combo-metadata.test.ts
• node --import tsx/esm --test tests/unit/resolveComboContextLimit.test.ts
• GET /v1/models for a combo whose catalogued members are 500k (plus an unsourced member) advertises 500000, not 128000

[diegosouzapw]

Verified this fix end-to-end -- the production change is correct and does what the title says.

Confirmation: I built a real (not fabricated) combo {glm/glm-5.2, aimlapi/deepseek-chat} -- aimlapi/deepseek-chat is a genuinely registered model whose resolveTokenLimit() falls through to the generic 128k catch-all (specific: false), matching the exact scenario in #10734. With this PR's getSourcedTokenLimit() in place, computeComboContextLength() returns 1000000 as expected. Reverting just the two getSourcedTokenLimit call sites back to getTokenLimit reproduces the bug exactly (128000). All 4 tests named in the test plan pass, plus the shipped combo-context-generic-default-10734.test.ts suite (7/7).

One suggestion (non-blocking): 3 of the 7 shipped tests -- including the flagship "GET /v1/models does not advertise 128k" integration test -- use not-a-real-provider/no-such-model as the "unsourced member" fixture. That model is filtered out by a separate, pre-existing gate (getCanonicalModelMetadata's source check for "not registered in any source at all"), independent of this fix. I confirmed this by reverting the production change and re-running the suite: all 7 tests still pass unchanged. Only the 4 lower-level resolveTokenLimit/getSourcedTokenLimit unit tests actually fail without the fix.

For a regression guard that actually exercises getComboTargetCatalogMetadata/computeComboContextLength's new branch, consider swapping that fixture for a real registered-but-unsourced model (e.g. aimlapi/deepseek-chat, or any PROVIDER_MODELS entry with no canonical contextWindow and no REGISTRY.defaultContextLength/DEFAULT_LIMITS entry for its provider) in a follow-up. Doesn't block this PR -- the production fix is sound and well-scoped as-is.

⚠️ Noted: base-red #9985 is correctly flagged in the PR body.