#2036 [Linux] --calibrate combined with --model silently falls through to plain re-setup (setup.py:4958)

Description

## Problem

`./setup.sh --calibrate --model IQ3_XXS --yes` does **not** calibrate. `setup.py:4958`:

```python
if have and a.calibrate and not (a.setup or a.model or a.family or a.check):
```

The calibrate branch is skipped whenever `--model` is also present; execution falls through to the plain setup path, which re-runs setup and then idles waiting for a browser. No warning is printed that `--calibrate` was ignored.

## Reproduction (Linux, RTX 3060 12G, build `7695859`)

1. `./setup.sh --calibrate --model IQ3_XXS --yes </dev/null` -> no calibration runs; setup re-runs (and silently rewrites the engine config, see #388) and idles with a validation server on :8080.
2. `./setup.sh --calibrate --yes </dev/null` (no `--model`) -> calibration runs normally (15-45 min), writes tuned args into the model config, then serves the model on :8080 as its own validation step - that idle IS the finished state.

## Impact

The natural invocation (`--calibrate --model <name>`, "calibrate this specific model") silently does a different thing: a full re-setup that also rewrites engine config. In an unattended/systemd bootstrap this burns 10+ minutes, occupies the serving port, and the operator believes calibration happened. It cost a full calibration cycle during our switchover before we found the guard clause.

## Expected

Either accept `--model` as the calibration target, or fail fast with an explicit message: `--calibrate cannot be combined with --model; run bare --calibrate`.

## Acceptance criteria

- `--calibrate --model X` either calibrates model X or exits with a clear error.
- No silent fall-through to plain setup when `--calibrate` was requested.