Agent Host: first-turn config refresh leaves new chat permanently active

- Copilot Chat Extension Version: N/A (built-in Agent Host)
- VS Code Version: 1.134.0-insider
- OS Version: macOS 26.6.1 arm64
- Feature (e.g. agent/edit/ask mode): Agent mode, new worktree-isolated session
- Selected model (e.g. GPT 4.1, Claude 3.7 Sonnet): GPT-5.6 Sol
- Logs: Agent Host debug bundle captured locally. The bundle did not include the Copilot runtime process log containing the terminal failure, so the relevant redacted excerpt from `process-1786891402662-64463.log` is included below. Affected Copilot CLI runtime: 1.0.79.

Steps to Reproduce:

1. Create a new worktree-isolated Agent Host session with active client tools and synced customizations.
2. Send the first prompt so the provisional chat materializes while the worktree-scoped customization bundle is republished.
3. Observe that the chat remains active indefinitely with no assistant response or terminal error.

## Expected

The first turn completes normally, or a runtime failure produces a terminal `chat/error`/`chat/turnComplete` transition so the UI stops showing the chat as active.

## Actual

The turn remains permanently active. In the captured run:

- Host session: `copilotcli:/98345434-f968-419e-bc34-dd191b28e382`
- SDK session: `265899d1-d87c-4303-b143-26355624c196`
- Turn: `request_dec9bbae-59c0-463b-99d4-015c2a95d36b`

The AHP turn opened optimistically at `20:24:57.249Z`, before an SDK session existed. The provisional SDK session was created at `20:25:01.105Z`. Its worktree-scoped synced-customization bundle was then republished with a new content nonce while the client tool set remained unchanged. During the same first-send call, `ActiveClient.requiresRestart` detected the updated plugin snapshot and refreshed the session.

Agent Host log:

```text
Session configuration changed, refreshing session
SDK resumeSession failed: Session not found: 265899d1-d87c-4303-b143-26355624c196
Resume reported no session history; falling back to createSession with same ID
Fallback createSession succeeded
session.send() returned
```

Copilot runtime process log:

```text
20:25:01.105Z Created session: 265899d1-d87c-4303-b143-26355624c196
20:25:01.364Z Closing session 265899d1-d87c-4303-b143-26355624c196
20:25:01.372Z Destroyed session 265899d1-d87c-4303-b143-26355624c196 after last owner detached
20:25:01.380Z Session not found: 265899d1-d87c-4303-b143-26355624c196
20:25:01.403Z Created session: 265899d1-d87c-4303-b143-26355624c196
20:25:11.338Z Agentic loop failed: Error: CompactionProcessorHandle has already been disposed
```

No `session.error`, `session.idle`, assistant turn event, or matching AHP `chat/turnComplete` reached Agent Host afterward. The only chat actions after `chat/turnStarted` were worktree-creation activity and one response-part placeholder.

## Root cause

The primary trigger is deterministic in VS Code:

1. The first send materializes a provisional SDK session.
2. The worktree-scoped customization snapshot changes before that same send reaches the model.
3. `_sendMessageOnce` destroys the just-created empty session and attempts to resume the same SDK ID.
4. Disconnecting the last owner destroys that empty runtime session, so resume reports `Session not found`.
5. `CopilotSessionLauncher` falls back to `createSession` with the same ID.

The recreated runtime then reports that its `CompactionProcessorHandle` is disposed. The logs prove the failure follows same-ID destroy/recreate, but do not prove the precise internal mechanism. The replacement may be observing state retained from the first instance, or late teardown from the first instance may be disposing a resource used by the replacement.

The runtime catches the agentic-loop failure but no terminal SDK event reaches Agent Host, leaving the optimistic AHP turn active.

## Existing hang detection

`AgentHostTurnTracker` already classifies a turn with no observed provider progress as `noProgress` after `TURN_HANG_THRESHOLD_MS` (five minutes), but this watchdog only emits telemetry. It does not terminate or recover the active turn.

## Potential fixes

1. Avoid destroy→resume for a newly materialized, unused session during its first send. Reconcile the final worktree-scoped customization snapshot before creating the SDK session, or recreate the empty backing without reusing the destroyed runtime identity.
2. Use the existing `AgentHostTurnTracker` no-progress watchdog as a recovery hook so a lost provider turn cannot remain active indefinitely.
3. Harden same-ID session recreation in the Copilot runtime and ensure agentic-loop failures always produce terminal SDK events.

Unrelated MCP startup/auth warnings occurred concurrently but were not the source of the disposed-handle failure.