[Feat] Add Mind reader mode for expanded LLM thoughts

[roomote-roomote]

Created by Roomote. Follow up by mentioning @roomote-roomote, in the web UI, or in Slack.

What changed

• Added a user-level Mind reader mode preference to Personal Settings, immediately above Narration mode.
• Persisted the preference in user metadata with a default of disabled.
• Merged preference metadata atomically so concurrent setting changes cannot overwrite each other.
• Applied the preference as the default expansion state for LLM thought blocks while preserving manual conversation-level expand and collapse choices.
• Added coverage for persistence, concurrent updates, optimistic preference updates, settings interaction, and reasoning expansion behavior.

Why this change was made

Users who prefer to follow model reasoning can now opt into seeing LLM thoughts expanded automatically without reopening each block.

Impact

Mind reader mode is off by default, so existing conversation behavior is unchanged until a user enables it. Once enabled, thought blocks open automatically by default, while manual choices within a conversation continue to take precedence.

[roomote-community]

No new code issues found. See task

• apps/web/src/trpc/commands/preferences/index.ts:120 The new preference is saved by separate mutations, but each request reads and later replaces the whole metadata object. If a user enables Mind reader mode and Narration mode before either request finishes, both reads can see the same old metadata and the later update erases the other preference. Merge the JSON keys atomically (or serialize these writes) and cover concurrent updates.

_{Reviewed 7a5d175}

[roomote-roomote]

Fixed the concurrent preference-write race in 7a5d175. Preference metadata now uses an atomic JSONB merge, and a database regression test verifies simultaneous Mind reader and Narration updates are both preserved.