[Improve] Gate the linked-issue fetch on the routing precheck; default routing to Gemini 3.6 Flash

[daniel-lxs]

What changed

Follow-up to #693.

Two-step gate. gatherContextFromConfiguredMcps now runs a tool-free routing precheck first and only calls gatherExternalIssueContext when the precheck returns needsExternalLookup=true — finally wiring the signal that has been telemetry-only since the old two-step design. When context is fetched, routing runs once more with the issue as untrusted reference material and reports phase: 'mcp' with the tools used. Fail-open is unchanged: an unavailable integration, inaccessible issue, or empty fetch keeps the precheck decision (phase: 'direct').

Routing model default. Workspace routing (Slack/Linear/Discord/etc. and GitHub paths) now resolves its model as context.routingModel → new R_ROUTER_MODEL deployment override → google/gemini-3.6-flash, instead of falling through to the deployment-wide small model. The decision's debug model field now reports the actual resolved model instead of the roomote-small-model placeholder.

Why

#693 paid the issue-fetch deadline (up to 8s) on every routing call containing a pasted GitHub/Linear issue link, even when the message alone already routed. In the router external-lookup gate eval (24 labeled cases × 4 repetitions × 5 candidate models, run against the production routing prompt), roughly two-thirds of link-bearing messages routed correctly with no fetch — and Gemini 3.6 Flash was the only stable gate: 0% false lookups and 100% lookup recall in every repetition, at ~4.5s p50 per call. Fetched issue context took genuinely ambiguous links from ~50% blind accuracy to 100% for every model, so the fetch stays — it just only runs when it can change the decision.

Trade-off: when a lookup is needed, routing now costs two model calls plus the fetch (one extra call vs #693). The eval says that's the minority of link-bearing tasks.

Tests

• Reworked the [Improve] Route tasks using linked issue context #693 routing test for the gated flow (precheck asks → fetch → informed re-route, phase: 'mcp').
• New: precheck routes without external context → no MCP call, single model call, phase: 'direct'.
• New: precheck asks but fetch fails → precheck decision stands, fail-open.
• vitest run src/server/router: 12 files, 94 tests green; slack router-debug/auto-route consumers green; tsc --noEmit clean.

🤖 Generated with Claude Code

[roomote-roomote]

1 issue outstanding. See task

• packages/cloud-agents/src/server/router/types.ts:29 The hard-coded google/gemini-3.6-flash bypasses the deployment-selected model and is passed directly to the OpenCode SDK. Deployments configured with OpenRouter, Bedrock, etc. but no GEMINI_API_KEY will now fail every Slack/Linear/Discord/GitHub routing call and fall back. Resolve this default through the configured provider or retain the deployment small-model fallback when no router override is set.

_{Reviewed 65087fd}