Final review fixes: QUICKSTART weights note, REPRODUCE prereq, export-script docstrings

- QUICKSTART.md: drop stale 'no trained weights ship' line; document the
  bundled checkpoints/atome_1m_v1.pt path alongside the train-from-scratch
  flow so first-time cloners aren't misled.
- REPRODUCE.md: add explicit scripts/build_corpus.py --source tinystories
  prerequisite for the 944K training commands that read
  data/tinystories_full.txt.
- scripts/export_to_atome_packed.py and export_to_atome_p3.py: lead the
  docstring with **Experimental** + a pointer to FRONTIER.md, so it's
  visible at first glance that these formats don't have a C decoder yet.

Author: Atome LM Team

QUICKSTART.md

@@ -64,6 +64,16 @@ is absent.
 
 ---
 
-**No trained weights ship with this repository** — Atome LM is
-train-from-scratch. Pre-trained 944 K-parameter checkpoints, MCU firmware,
-and full benchmarks are part of the commercial kit (see [README.md](README.md)).
+**Trained weights are bundled** in `checkpoints/` — `atome_944k.bin`
+(packed C-engine blob), `atome_1m_v1.pt` (PyTorch source), and
+`vanilla_1m_v1.pt` (FP32 baseline for the 944 K reversal A/B in
+[HONEST_RESULTS.md](HONEST_RESULTS.md)). If you want to run the model
+without training first:
+
+```bash
+python scripts/demo.py --checkpoint checkpoints/atome_1m_v1.pt
+```
+
+If you want to train your own from scratch, follow the
+`scripts/train_demo.py` flow above — it produces a 60 K-parameter model
+in ~30 min on a CPU.

REPRODUCE.md

@@ -32,6 +32,15 @@ python3 -c "import json; r=json.load(open('ab_results.json')); \
 
 The ckpt that produced val_loss 1.0545 / ppl 2.87.
 
+Prerequisite — build the full TinyStories corpus (one-shot, ~5 min):
+
+```bash
+PYTHONPATH=. python3 scripts/build_corpus.py --source tinystories \
+    --output data/tinystories_full.txt
+```
+
+Then train:
+
 ```bash
 # On the GPU pod:
 PYTHONPATH=. python3 scripts/train_atome_1m.py \

scripts/export_to_atome_p3.py

@@ -1,6 +1,11 @@
 #!/usr/bin/env python3
 """scripts/export_to_atome_p3.py — write a power-of-3 ATOME-P3 binary.
 
+**Experimental.** The C engine in `c_engine/upstream/` does not yet decode
+ATOMEP3; this script produces research artifacts only. For an
+MCU-deployable blob today, use `scripts/export_to_atome.py` (ATOME01).
+See `FRONTIER.md` for the ATOMEP3 / ATOME02 status.
+
 Identical wire layout to ATOME01 except every quantized payload stores
 3-bit power-of-3 codes (3 bits / weight, 8 codes per 3 bytes), not
 2-bit ternary trits (2 bits / trit, 4 trits per byte).

scripts/export_to_atome_packed.py

@@ -1,6 +1,11 @@
 #!/usr/bin/env python3
 """scripts/export_to_atome_packed.py — write a base-3 packed `ATOME02` binary.
 
+**Experimental.** The C engine in `c_engine/upstream/` does not yet decode
+ATOME02; this script produces research artifacts only. For an MCU-deployable
+blob today, use `scripts/export_to_atome.py` (ATOME01). See `FRONTIER.md`
+for the ATOME02 / ATOMEP3 status.
+
 Identical wire layout to `export_to_atome.py` (ATOME01) except every
 "packed trits" payload is laid out as