fix(core): normalize RPC handler failures

[kitlangton]

Why

An unexpected RPC handler failure bypasses ordinary Effect error recovery for in-process callers, while HTTP callers receive rpc.internal. The same call should have a recoverable failure at either entry point.

What Changes

• Before: thrown exceptions, defects, undeclared failures, and invalid declared-error payloads escape direct calls as defects. HTTP defect responses can expose the exception message.
• After: handler defects are logged and become typed { type: "rpc.internal", message: "RPC call failed" } failures. HTTP maps them to 500; its fallback also logs rather than echoes defects outside handler execution.

The existing declared-error encoder stays in place. Valid declared errors, successful results, input/output validation, and interruption retain their behavior.

Same caller, different outcome

For example, inside a scoped Effect.gen, with rpc obtained from Rpc.Service:

const Weather = Rpc.define({
  id: "weather",
  methods: {
    forecast: { input: Schema.String, output: Schema.String },
  },
  events: {},
})

yield* rpc.register(Weather, {
  forecast: () => Effect.die(new Error("backend unavailable")),
})

// The caller is identical before and after this PR.
const result = yield* rpc.client(Weather).forecast("Boston").pipe(
  Effect.catchIf(
    (error) => "type" in error && error.type === "rpc.internal",
    () => Effect.succeed("Weather unavailable"),
  ),
)

// BEFORE: the defect escapes; catchIf never runs and no result is produced.
// AFTER:  result === "Weather unavailable"; the original defect is logged.

The failed handler does not become successful. It becomes an ordinary RPC failure that the caller can choose to recover from, instead of requiring a separate Effect.catchDefect at each call site.

HTTP response for the same handler failure

// BEFORE: HTTP 500, with the handler's exception message exposed.
{
  "_tag": "RpcInternalError",
  "type": "rpc.internal",
  "message": "backend unavailable"
}

// AFTER: HTTP 500, with a generic message; details remain in server logs.
{
  "_tag": "RpcInternalError",
  "type": "rpc.internal",
  "message": "RPC call failed"
}

Scope

Supersedes #46902 with a smaller implementation and only seven regression cases, not the restored RPC test suite. Production changes are confined to core/src/rpc.ts and server/src/handlers/rpc.ts.

All seven tests were run against unmodified v2 at c992716523 and failed for the intended reason:

Cases	Before	After
Five direct-call cases: defect, throw, raw failure, undeclared error, invalid error payload	Defect bypasses recovery	Typed rpc.internal, logged once
HTTP handler defect	Raw handler message returned	Generic HTTP 500
HTTP input-schema defect	Raw schema message returned	Generic HTTP 500

Unlike #46902, undeclared-error names also stay in logs rather than the public message.

Verification

# packages/core
bun typecheck
bun run test test/rpc-handler-errors.test.ts --rerun-each 5
bun run test

# packages/server
bun typecheck
bun run ../core/script/test.ts test/rpc-handler-errors.test.ts --rerun-each 5
bun run ../core/script/test.ts

• Both package typechecks pass.
• Before the fix: 5 Core + 2 HTTP failures. After: 35/35 repeated regression tests pass.
• Full Core suite: 4022 pass, 39 skip, 0 fail.
• Full Server suite: 52 pass, 3 skip, 0 fail.
• A separate runtime probe confirmed that valid declared errors, successful values, invalid-input/output errors, and cancellation remain unchanged.