This checkpoint ships **uncalibrated per-tensor activation scales**. This PR documents that in the model card and adds a notice to `.quant_summary.txt`. It does **not** fix the weights.
## What was verified
- Every `*_input_scale` tensor in the published safetensors is exactly `1.0` — **57,600 of 57,600**, no exceptions (75 quantized layers x 256 experts x 3 projections).
- `.quant_summary.txt` contains 150 calibrated `*_input_quantizer` entries (75 layers x {`gate_up_proj`, `down_proj`}) with real `amax` values, so calibration **was** run — the values just never reached the export.
- Expected scale is `input_scale = amax / (6 * 448) = amax / 2688`:
| tensor | `amax` | expected | published |
| --- | --- | --- | --- |
| `model.layers.3.mlp.experts.gate_up_proj` | 4.10e-01 | 1.5253e-04 | **1.0** |
| `model.layers.3.mlp.experts.down_proj` | 1.00e+00 | 3.7202e-04 | **1.0** |
- The weight side is unaffected and confirms the formula: `gate_up_proj_weight_quantizers.0` `amax=2.12e…
