Update .NET Copilot SDK cookbook for current SDK releases (#3548)

* Update .NET Copilot SDK cookbook: fix On<T> breaking change, add in-process runtime, tool search, and message provenance recipes

- Fix breaking change: Session.On is now generic (On<T> where T : SessionEvent).
  Updated all recipes that used the old non-generic session.On(evt => ...) pattern:
  error-handling, pr-visualization, managing-local-files, accessibility-report,
  ralph-loop (both .md and .cs where applicable).
- Add new 'In-Process Runtime' recipe (in-process-runtime.md + recipe/in-process-runtime.cs)
  demonstrating RuntimeConnection.ForInProcess(), live-tested.
- Document tool search / deferred loading (SessionConfig.ToolSearch / ToolSearchConfig)
  in pr-visualization.md, live-tested.
- Document typed message provenance (MessageOptions.Source / MessageSource) in
  error-handling.md, live-tested.
- Fix dotnet/README.md recipe index: it was missing Ralph Loop and Accessibility
  Report entries; also added the new In-Process Runtime entry.
- Add In-Process Runtime entry to the top-level cookbook/copilot-sdk/README.md
  .NET section.
- Update dotnet/recipe/README.md runnable-examples table with the new recipe.

All existing .NET cookbook recipes were rebuilt and rechecked against the current
GitHub Copilot SDK (Copilot CLI 1.0.85); no other breaking changes were found.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Update cookbook model names to current supported models

- Replace stale/nonexistent model IDs (gpt-5, claude-sonnet-4.5,
  claude-opus-4.6, gpt-5.1-codex-mini) with current values from the
  supported AI models list
- Prefer Model = "auto" wherever a specific model isn't load-bearing
  to the recipe's teaching point
- ralph-loop: pin gpt-5.3-codex (current codex-family GA model), since
  a dedicated coding model is a deliberate, documented choice for the
  autonomous loop
- multiple-sessions: reshape the model narrative so it demonstrates a
  real explicit-override use case (A/B testing via claude-sonnet-5)
  instead of two sessions sharing the same hardcoded model

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

* Use SendAndWaitAsync in in-process runtime recipe

Replace the event-based TaskCompletionSource pattern with SendAndWaitAsync so the example waits for the final assistant response, propagates session errors, and uses the SDK timeout behavior.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

---------

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

cookbook/copilot-sdk/dotnet/pr-visualization.md

@@ -164,7 +164,7 @@ await client.StartAsync();
 
 var session = await client.CreateSessionAsync(new SessionConfig
 {
-    Model = "gpt-5",
+    Model = "auto",
     OnPermissionRequest = PermissionHandler.ApproveAll,
     SystemMessage = new SystemMessageConfig
     {
@@ -185,7 +185,7 @@ The current working directory is: {Environment.CurrentDirectory}
 });
 
 // Set up event handling
-session.On(evt =>
+session.On<SessionEvent>(evt =>
 {
     switch (evt)
     {
@@ -256,3 +256,37 @@ while (true)
 | Flexibility     | Fixed logic       | **AI decides best approach**      |
 | Chart types     | What you coded    | **Any type Copilot can generate** |
 | Data grouping   | Hardcoded buckets | **Intelligent grouping**          |
+
+## Deferring tools with tool search
+
+The GitHub MCP Server alone exposes dozens of tools, and this recipe also has the file and code
+execution tools available — well past the point where stuffing every tool description into the
+model's context on every turn helps more than it costs. `SessionConfig.ToolSearch` controls when
+the SDK defers less-frequently-needed tools behind a search step instead of listing them all
+up front:
+
+```csharp
+var session = await client.CreateSessionAsync(new SessionConfig
+{
+    Model = "auto",
+    OnPermissionRequest = PermissionHandler.ApproveAll,
+    ToolSearch = new ToolSearchConfig
+    {
+        Enabled = true,
+        DeferThreshold = 20
+    },
+    SystemMessage = new SystemMessageConfig { Content = "..." }
+});
+```
+
+- `Enabled` turns tool search on or off explicitly; leave it `null` to use the runtime default.
+- `DeferThreshold` is the tool count above which MCP/external tools are deferred behind a
+  `tool_search_tool` call instead of being listed directly (the runtime default is 30 when
+  unset). Lowering it — as above — is useful once a recipe pulls in a large MCP server like the
+  GitHub MCP Server alongside file and code-execution tools, since it keeps the per-turn tool
+  list small while still letting Copilot search for the exact tool it needs.
+- To customize how the search itself behaves, register your own tool named `"tool_search_tool"`
+  with `OverridesBuiltInTool = true` to replace the built-in implementation.
+
+This is a per-session setting, not a global one, so you can tune it independently for
+tool-heavy recipes like this one versus lighter sessions elsewhere in your app.