[Fix] Fast personalization changes when users update preferences mid-conversation

[roomote-roomote]

​Opened on behalf of @mrubens. Follow up by mentioning @roomote-roomote, in the web UI, or in Slack.

What changed

• Snapshot each participant's personalization the first time they speak in a Fast conversation and reuse that participant-specific snapshot on later turns.
• Keep direct instructions effective through the current message and native conversation history while conversational learning continues to update the user's profile for future conversations under the frozen learning setting.
• Verify the exact conversation-and-user snapshot before saving a Fast preference, while standard task updates continue honoring the current profile setting.
• Document that settings edits and resets affect new conversations, or participants who have not spoken in the current conversation yet.

Why this change was made

This is a follow-up to #2482. Reloading personalization on every Fast turn changed the system prompt after learned preferences or settings edits, invalidating the provider's reusable prompt prefix. Stable per-participant snapshots preserve that cacheable prefix while retaining current-speaker adaptation and original-requester precedence.

Impact

Fast conversations now keep each speaker's system-prompt personalization stable, including when a shared thread switches away and back to that speaker. New conversations receive the latest saved profile.

Snapshots are stored in a dedicated encrypted table keyed by conversation and user. They are not added to shared conversation, Session, task, or transcript metadata; exact participant scoping prevents one speaker from receiving another speaker's private note, and canonical and compatibility transcript writes remain scrubbed. Fast preference writes use that persisted pair as the authority rather than trusting a caller-provided learning flag. Standard task start/resume behavior is unchanged because it already assembles private context at the run boundary and continues applying live opt-out.

Focused database, Fast service, shared prompt, and standard-task personalization tests pass. Repository lint, fast type checks, Knip, docs validation, broken-link checks, and diff checks also pass.

[roomote-community]

No code issues found. See task

• packages/db/src/lib/user-personalization.ts:328 Mid-conversation learning toggles disagree with the frozen snapshot, so a current speaker's new preference can be rejected instead of being saved for future conversations.

_{Reviewed 94bdf59}