fix(resilience): honor shared-registry passthrough providers (#11071)

[yourspraveen]

Ports the fix from #11075 (merged to main as c68cda7df) onto release/v3.8.50, where it never landed. Closes the part of #11071 that is still open on this branch.

Why this isn't a duplicate of #11078

#11078 fixed #11071 here with a different mechanism — isLocalProviderId(canonicalId) || isSelfHostedChatProvider(canonicalId). That covers local backends. It does not cover the shared registry.

hasPerModelQuota() currently consults the open-sse REGISTRY passthrough set, isCompatibleProvider, and the local/self-hosted families. None of them reads src/shared/constants/providers/, where 111 providers declare passthroughModels: true. Measured against the current tip:

total providers:                    348
declare passthroughModels: true:    111
NOT covered on release/v3.8.50:      40   ← novita, uncloseai, orcarouter, cheaperinference,
                                            freebuff, piapi, getgoapi, laozhang, thebai, …

For those 40, a missing-model 404 still cools the entire connection instead of locking out the one model — the original #11071 defect, still live for non-local providers. On a gateway where each upstream carries a different model subset, a single missing model takes the whole connection offline.

The change is one lookup, placed beside the existing passthrough check:

if (getProviderById(canonicalId)?.passthroughModels === true) return true;

The test targets the residual gap, not #11075's original cases

Worth flagging for review, because a verbatim port would have been misleading:

• fix(resilience): honor shared passthrough providers #11075's case list (ollama-local, lm-studio, vllm) already passes here via fix(resilience): keep Ollama model-not-found failures scoped to connection model lockout (#11071) #11078's local branch — it would prove nothing on this branch. It is still asserted, to guard against regressing fix(resilience): keep Ollama model-not-found failures scoped to connection model lockout (#11071) #11078.
• novita / uncloseai / orcarouter are the real cases: they fail before this change, pass after.
• fix(resilience): honor shared passthrough providers #11075's negative control (mlx-gemma => false) is deliberately not ported. It declares no passthrough metadata but is local, so fix(resilience): keep Ollama model-not-found failures scoped to connection model lockout (#11071) #11078 correctly makes it true here — porting that assertion would have contradicted fix(resilience): keep Ollama model-not-found failures scoped to connection model lockout (#11071) #11078. openai / anthropic serve as the negative control instead.

Verification

Suite	Result
account-fallback-service.test.ts	91/91
25 lockout / cooldown / passthrough suites	317/318
typecheck:core, lint (with suppressions), check:mutation-test-coverage	clean

The single failure is quota classifier rejects terminal-looking evidence on ineligible statuses in combo-quota-exhaustion-only-fallback.test.ts — an inherited base-red that fails identically on the untouched tip (also red on #11146 and #11154).

⚠️ base-red inherited: #9985

Note on the diff

Four hunks in account-fallback-service.test.ts outside the new test are formatter output, not intentional edits — those lines exceed the 100-char width and the file was not prettier-clean on the base, so lint-staged rewraps them on any commit touching it.