Fix reasoning-only replies and TODO continuation loops

[salacoste]

Summary

This PR fixes two user-facing regressions observed in OpenCode sessions:

1. Responses that only contain reasoning/tool output resulted in no visible final answer.
2. Sessions with pending TODOs could loop indefinitely even when the user did not request continuation.

What changed

1) Final-response fallback for reasoning/tool-only replies

• Detects assistant messages that finish without visible text but contain reasoning/tool parts.
• Enqueues a synthetic user message and routes it through a hidden final agent with tools disabled.
• Ensures a plain-text final response is returned.

2) TODO continuation loop guard

• Adds todo_state tracking (paused/updatedAt/lastUpdatedMessageID).
• Pauses automatic TODO continuation unless the user explicitly asks to continue.
• Clears pause on todowrite updates and injects a short system guard when paused.

3) Supporting updates

• Added documentation for both fixes.
• Adjusted a Bun test to avoid test.concurrent (not supported by Bun).

Files touched (high level)

• packages/opencode/src/session/prompt.ts
• packages/opencode/src/session/llm.ts
• packages/opencode/src/session/todo.ts
• packages/opencode/src/tool/todo.ts
• packages/opencode/src/agent/agent.ts
• packages/opencode/src/agent/prompt/final.txt
• packages/opencode/test/session/retry.test.ts
• docs/changes-2026-01-07.md
• docs/reasoning-only-fallback-plan.md
• docs/todo-loop-guard.md

Tests

• bun test (packages/opencode)
• bun turbo typecheck (pre-push hook)

Notes

• The TODO guard is server-side; UI changes are not required.
• The fallback uses a hidden final agent with tools disabled to avoid recursive tool calls.