fix(session): don't clobber user's model swap on model-less prompts

[Qiiks]

Issue for this PR

Closes #42893

Type of change

• Bug fix

What does this PR do?

A model-less prompt (no input.model — e.g. plugin completion-reminders via promptAsync) resolves from the agent fallback (ag.model) for the turn. The existing design (PR #26765) intentionally uses the agent's model when no explicit model is given — this is correct for the turn's model and is preserved.

However, setAgentModel then overwrites the session row to the resolved model, permanently destroying a model the user explicitly stored via a prior swap. The TUI is unaffected (always sends draft.model explicitly), but opencode serve / mobile clients lose the swap on the next plugin reminder.

The fix: guard the setAgentModel write in createUserMessage — skip it when:

1. The prompt carried no explicit model (input.model == null), AND
2. The session already has a stored model that differs from the agent's resolved model (storedModelDiffers)

Fresh sessions (no stored model) still persist the agent's model on the first model-less prompt (hasNoStoredModel guard), and explicit-model prompts always persist normally. The model-resolution precedence is unchanged — model-less prompts still use the agent's model for the turn.

How did you verify your code works?

• New regression test: "model-less prompt does not clobber user's stored model swap" — creates a session, swaps to kimi-k2.5-free, sends a model-less prompt with an agent carrying test/test-model, asserts the turn uses test-model (design preserved) but the session row stays kimi-k2.5-free (the fix). Verified RED→GREEN.
• Fresh-session test: "model-less first prompt on fresh session persists agent model to row" — confirms the hasNoStoredModel guard fires setAgentModel on a fresh session so the row is initialized.
• Existing test preserved: "applies agent variant only when using agent model" (the test that pins the agent-model design from Persist session model switches outside event flag #26765) still passes — model-less prompts still use the agent's model for the turn.
• Full suite: 46 pass / 14 skip / 0 fail across test/session/prompt.test.ts.
• Typecheck: 30/30 packages clean (turbo typecheck).

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.