Pitch

This was filed and deferred (#315). Re-opening as a v0.8.6 deliverable because it's specifically a DeepSeek-TUI superpower nobody else can claim — we know V4's two-tier shape better than any provider-neutral tool.

`/model auto` flips routing to a heuristic: each turn picks Pro or Flash based on signals:

• Flash (cheap): tool-only turns (file reads, git inspection, small edits), follow-ups inside an established context, status / summary requests.
• Pro (expensive): explicit reasoning requests ("think about"), debugging, architecture, security review, anything tagged with `/effort high` or `/effort max`.

Heuristic visible to the user: a small chip in the footer reads `auto · flash` or `auto · pro` so they can see what's being chosen and override at any time.

Implementation sketch

• New `AutoModelRouter` in `crates/tui/src/core/auto_router.rs`.
• Inputs: turn-N message text, recent tool history, current effort level, last few turns' usage.
• Output: `ModelChoice::Pro | ModelChoice::Flash`.
• Wire into engine's `request_model` instead of `session.model` when `auto` is on.
• Per-turn override: `Ctrl+M` toggles between auto/pro/flash for the next send.

Acceptance

• `/model auto` enables auto-routing; `/model deepseek-v4-pro` re-pins.
• Footer chip shows the resolved model per turn.
• `Ctrl+M` per-send override.
• Heuristic test fixtures cover all six combinations.