[discuss] Per-view tRPC instances for DocumentDB webviews (remove ctx casts)

Start with src/webviews/_integration/trpc.ts, then inspect the affected view routers and controllers that use the shared instance and context casts. Compare the per-view integration shape referenced from vscode-cosmosdb. This discussion is not ready for implementation until the instance strategy and actionContext modelling are decided; done would be a recorded decision and consistent treatment across the affected views.

Context

@microsoft/vscode-ext-webview supports either one shared tRPC instance for all webviews or one instance per webview (initWebviewTrpc<ViewContext>()). The DocumentDB extension currently uses a single shared instance bound to the framework BaseRouterContext:

// src/webviews/_integration/trpc.ts
const trpc = initWebviewTrpc<FrameworkBaseRouterContext>();

Because that one instance serves CollectionView, DocumentView, QueryInsights, IndexView, etc., two consumer-side patterns fall out of the design (surfaced during review of #795):

1. Per-procedure narrowing cast — every procedure re-narrows ctx to its view type:

   const myCtx = ctx as RouterContext; // ~30 sites
   

2. Omit on the root context — controllers build the initial context, but actionContext is injected per-call by the telemetry runner, so the root object can't be a full RouterContext:

   const trpcContext: Omit<RouterContext, 'actionContext'> = { ... };
   

Neither is a package limitation — both are how this extension wires the package. vscode-cosmosdb avoids the cast by building a separate instance per webview, each initTRPC.context<QueryEditorRouterContext>().

Proposal

Move DocumentDB to per-view tRPC instances — one initWebviewTrpc<ThisViewContext>() per webview — so ctx is already the precise per-view type inside procedures.

Pros

• Removes ~30 ctx as RouterContext casts. Procedures read ctx.sessionId, ctx.actionContext, etc. directly, fully typed.
• Type safety instead of casts. A wrong field access becomes a compile error rather than being masked by the as cast.
• Matches the framework's intended usage and the vscode-cosmosdb shape, making the two extensions' integration layers converge (easier shared learnings / future extraction).
• Clearer per-view boundaries — each view owns its context type end to end; no giant shared BaseRouterContext union of concerns.

Cons

• More wiring / boilerplate. N instances instead of one: each needs its own publicProcedure / router / createCallerFactory (and the telemetry middleware applied per instance). The framework note warns against hiding the .use() chain behind a generic helper (it collapses tRPC's ProcedureBuilder inference to any), so the wiring is repeated per instance by design.
• Circular-import care. The current single-leaf trpc.ts deliberately breaks an appRouter -> per-view router -> appRouter cycle. Per-view instances need the same discipline replicated N times.
• Does not remove the Omit (or optional) on the root context. actionContext is still injected per call, so the controller-built root context still lacks it regardless of instance count. This refactor fixes the cast, not the injected-field modelling.
• Churn for limited functional gain. It's an ergonomics/typing improvement, not a behavior change; touches every router + controller.

Decision needed

• Do we want per-view instances (remove casts) vs. keep the single instance (one cast pattern, less wiring)?
• Independently: how to model the injected actionContext on the root context — keep Omit<RouterContext, 'actionContext'>, make it optional, or a WithActionContext<T> wrapper at read sites?

Follow-up to #795 (kept out of that PR to stay scoped to the package API change).