docs: record the published v0.1.3 image served on H100s

docs/STATUS.md

@@ -29,7 +29,7 @@ and `ibm-granite/granite-switch-4.1-8b` unless a row names another checkpoint.
 | Tensor parallelism | the 8b with `--tensor-parallel-size 2`: the full battery | pass |
 | Long contexts | the 8b with `--max-model-len 131072`, `scripts/longctx_check.py` (9 checks) | 9/9: a 98k-token request answered in one engine read with the danger-zone header and rejected without it; the 131,072 backstop and the default 32k/64k caps enforced; `/v1/decisions` has no header escape |
 | Docker image build | `docker build --platform linux/amd64`, default and `WITH_ASR=1` variants | both build; the stack inside is package-identical to the live-tested one (PyPI vLLM 0.30.0, torch 2.13.0+cu130, transformers 5.17.0, granite-switch 0.1.1); entry points and imports verified in-container; the `WITH_ASR=1` image holds the pinned speech-model revision as its only snapshot, with `refs/main` pointing at it |
-| Serving from the image | the default image (no speech model inside) under enroot (no Docker daemon on the GPU nodes), through `serve.sh`, only the checkpoint mounted, Hub offline: the 8b on one H100 and the 30b on two H100s (`--tensor-parallel-size 2`, set from the GPU count). On each: `e2e_check`, `probe_check`, `accuracy_check` | pass on both: 11/11 and 6/6, accuracy as on the pip path; the log shows `hf_overrides: {'asr_enabled': False}` and never names the speech model |
+| Serving from the image | the default image (no speech model inside) under enroot (no Docker daemon on the GPU nodes), through `serve.sh`, only the checkpoint mounted, Hub offline: the 8b on one H100 and the 30b on two H100s (`--tensor-parallel-size 2`, set from the GPU count). On each: `e2e_check`, `probe_check`, `accuracy_check` | pass on both: 11/11 and 6/6, accuracy as on the pip path; the log shows `hf_overrides: {'asr_enabled': False}` and never names the speech model. Repeated with the published `ghcr.io/generative-computing/granite-switch-systemone:0.1.3` image (image manifest digest `sha256:77a56fb19768727de6e714d160d32842ddd1715964d51b9a5da5acba604e7971`, every layer's digest verified on download): same results on the 8b and the 30b, and in the speech-input row below |
 | Speech input on (`SYSTEMONE_SPEECH_INPUT=1`) | the 8b on the default image with the speech model supplied from a mounted Hub cache (`HF_HUB_CACHE`); and, as a negative control, the same switch with no speech model anywhere | the 8b starts and passes the same three checks; the control stops in `serve.sh` with the preflight message naming the model and the fixes, before vLLM starts. The `WITH_ASR=1` image (model baked in) was served on the 8b in an earlier run |
 | Decision Index runs | samples of the Decision Index suite (edition 0.3) through `/v1/systemone` at `--max-model-len 131072`, on the 8b and the 30b, as an end-to-end validation of the endpoint | completes with no errors; scores are not reported here while the methodology is verified, and the model card's published figures are the reference (see the caveats below) |