fix(tui): restore auto model mode on session load

[LING71671]

Summary

Closes #1797.

This fixes resumed sessions that were saved while /model auto was active. Before this change, loading such a session restored app.model to the literal string auto but left app.auto_model false in some resume/load paths. The next turn could then treat auto as a fixed provider model and send it directly to the DeepSeek API, which rejects it with HTTP 400.

Changes

• Added App::restore_model_selection so session restore paths rebuild the full model-mode state from persisted metadata.
• Updated both /load and startup/session-picker resume handling to use that helper.
• When the restored model is auto, the app now:
  • enables auto_model
  • clears stale effective-model labels
  • restores auto reasoning mode
  • recalculates the compaction budget against the default concrete budget model
• Added regression coverage for both explicit /load and apply_loaded_session resume paths.

Verification

• cargo fmt --all -- --check
• git diff --check
• cargo test -p deepseek-tui --bin deepseek-tui load_auto_model_session_restores_auto_mode
• cargo test -p deepseek-tui --bin deepseek-tui apply_loaded_session_restores_auto_model_mode
• cargo build
• cargo clippy --workspace --all-targets --all-features

Notes: build/clippy still report pre-existing warnings in schema_migration.rs for unused migration marker structs, and clippy reports two pre-existing needless_return warnings in tui/ui.rs. No new warnings were introduced by this change.

[gemini-code-assist]

Code Review

This pull request introduces a centralized restore_model_selection method in the App struct to handle the restoration of model settings when loading a session. This change ensures that the 'auto' model mode, reasoning effort, and compaction budget are correctly synchronized. Additionally, unit tests were added to verify that loading a session with an 'auto' model correctly restores the application state. I have no feedback to provide.

[LING71671]

Superseded by #1885, which carries the same #1797 fix forward on top of current main with the remaining /load path covered and fresh regression assertions. Closing this older dirty branch to keep the queue clearer.