feat(resilience): context overflow fallback, OAuth token detection, empty content guard & context-optimized combo strategy

[oyi77]

Summary

This PR introduces four resilience and intelligence improvements to the provider request pipeline: context overflow auto-fallback, OAuth token error classification, empty-content response detection, and a new `context-optimized` combo strategy.

---

Changes

1. Context Overflow Auto-Fallback

Files: `chatCore.ts`, `modelFamilyFallback.ts`, `errorClassifier.ts`

When a provider returns `400` with context-limit signals (e.g. "prompt too large", "token limit", "exceeds context"), the router now automatically falls back to the next model in the same family instead of propagating the error to the client.

• Added `isContextOverflowError(status, message)` in `modelFamilyFallback.ts` — 11 token-limit signal fragments with regex fallback
• Added `isContextOverflow(errorText)` in `errorClassifier.ts` — same signal list for classification
• Added `CONTEXT_OVERFLOW` to `PROVIDER_ERROR_TYPES`
• `chatCore.ts`: on overflow, calls `getNextFamilyFallback()`, retries with new model, logs `CONTEXT_OVERFLOW_FALLBACK`

2. OAuth Invalid Token Classification (T11)

Files: `accountFallback.ts`, `errorClassifier.ts`, `chatCore.ts`

Previously all `401`s not matching "account deactivated" were classified as generic `UNAUTHORIZED`, causing OAuth expiry to be treated as a hard failure.

• Added `OAUTH_INVALID_TOKEN_SIGNALS` + `isOAuthInvalidToken()` in `accountFallback.ts` — matches `"invalid authentication credentials"`, `"oauth 2"`, `"login cookie"`, `"invalid credentials"`
• Added `OAUTH_INVALID_TOKEN` to `PROVIDER_ERROR_TYPES`
• `classifyProviderError()` now checks OAuth signals before falling through on `401`
• `chatCore.ts`: `OAUTH_INVALID_TOKEN` logs a warning and updates connection state — node stays active since token refresh can recover

3. Empty Content Response Guard

Files: `errorClassifier.ts`, `chatCore.ts`

Some providers return `200 OK` with empty `content` — a silent fake-success that causes downstream failures.

• Added `isEmptyContentResponse(responseBody)` — handles OpenAI-style (`choices[].message/delta.content`), Anthropic-style (`content[]`), and raw `text`
• `chatCore.ts`: after response body is parsed, detects empty content, logs a `502`, triggers `getNextFamilyFallback()`, and on success recursively calls `handleChatCore` with the fallback model for a clean retry

4. Context-Optimized Combo Strategy

Files: `combo.ts`, `modelsDevSync.ts`, `contextManager.ts`, `modelFamilyFallback.ts`

New `context-optimized` strategy orders combo models by context window size (largest first), reducing overflow failures on large payloads.

• Added `getModelContextLimit(provider, modelId)` in `modelsDevSync.ts` — reads `limit_context` from synced models.dev capabilities
• Added `sortModelsByContextSize(models)` in `combo.ts`
• `handleComboChat()` now handles `strategy === "context-optimized"`
• `getTokenLimit()` in `contextManager.ts` now checks models.dev DB as priority 2 (after env override, before registry) — fixes gemini-2.5-pro: `1000000` → `1048576`
• Added `findLargerContextModel()` in `modelFamilyFallback.ts` for context-aware candidate selection

---

Test Updates

• `tests/unit/context-manager.test.mjs`: updated expected gemini-2.5-pro limit `1000000` → `1048576` to match models.dev synced value
• Minor import formatting cleanup

---

Test Plan

• Send a request exceeding a model's context window — verify `CONTEXT_OVERFLOW_FALLBACK` log and fallback model serves the response
• Simulate `401` with `"invalid authentication credentials"` — verify `OAUTH_INVALID_TOKEN` classification, node stays active
• Simulate `200` with empty `choices[].message.content` — verify `502` logged and fallback triggered
• Create a combo with `strategy: "context-optimized"` — verify models ordered largest context first
• Run `npm test` — context-manager unit test passes with updated gemini limit
• Verify `getTokenLimit("gemini", "gemini-2.5-pro")` returns `1048576`

🤖 Generated with Claude Code