kaggle(breeze,#412): pin the oracle's repetition penalty, and record the artifact readback

Both arms of the diff now DECLARE repetition_penalty = 1.0 on the greedy path
instead of both happening to inherit the same library default. The value does
not change — it is absent from generation_config.json, so HF's default already
applied — but an agreement that rests on a shared default is one that ends
silently the day the default moves, and this is the path whose whole purpose is
exact-integer equality.

Also records what the regenerated GGUFs actually contain, read back from the
published q4_k with a range request over its KV block rather than taken from
the converter's own log: theta 1e6 and eps 1e-6 (not the top-level decoys),
te.causal False, the norm fold flag the runtime demands, general.license
"other" — the tag that makes the NC gate fire — and the §4(b) notice verbatim.
778 tensors, 143 KV entries, q4_k 2.05 GiB.

Three decoys, dodged in the artifact, which is the only place it counts.

Refs #412

Author: CrispStrobe

docs/breeze-tts-2-port-notes.md

@@ -739,3 +739,28 @@ Still open (a validation item, not a known bug): *which* token history HF's
 backbone's `input_ids` are audio frames rather than text, so "the history" is
 ambiguous; the C++ applies it over the generated codebook-0 tokens. Confirm
 against a non-greedy reference run before treating the normal path as faithful.
+
+### Published artifacts, read back rather than asserted
+
+`cstr/breeze-tts-2-GGUF` regenerated 2026-09-15 with the norm fold:
+f16 5.32 GiB, q8_0 3.19 GiB, q4_k **2.05 GiB**, 778 tensors / 2 850.2 M live
+params, 351 tensors / 633.1 M dropped. The q4_k figure is what the quant
+policy predicts once `te.token_embd` is held at source precision; set
+`CRISPASR_BREEZE_QUANT_TEXT_EMBD=1` to include it and land near 1.75 GB.
+
+The KV block of the **published** q4_k was fetched with an HTTP range request
+and parsed (143 KV entries, 778 tensors). Verified in the shipped file, not in
+the converter's intentions:
+
+| key | value | what it rules out |
+|---|---|---|
+| `breeze.bb.rope_theta` | `1000000.0` | the top-level 500000 decoy |
+| `breeze.bb.rms_norm_eps` | `1e-06` | the top-level 1e-5 decoy |
+| `breeze.te.causal` | `False` | `use_bidirectional_attention: false` |
+| `breeze.te.rope_factor_full` | `8.0` | — |
+| `breeze.te.norm_weights_pre_offset` | `True` | a GGUF the runtime would reject |
+| `general.license` | `"other"` | a tag that would NOT trip the NC gate |
+| `breeze.license.notice` | verbatim, incl. `RESONIA, INC` | §4(b) |
+
+All three config decoys are therefore dodged *in the artifact*, which is the
+only place it counts.

tools/kaggle/breeze-refdump/breeze_refdump.py

@@ -150,7 +150,7 @@ def step(name, **kv):
 
 kh.init_progress()
 # Bump when the arms, capture or predicate change (see kh.provenance).
-SCRIPT_VERSION = "2026-09-15.6"
+SCRIPT_VERSION = "2026-09-15.7"
 # kh.provenance landed in the harness AFTER this kernel was first pushed, so the
 # 2026-09-02 run died in 10 s with AttributeError against its own fresh clone
 # (gotcha #24, the two-halves trap). Never let provenance logging be fatal.
@@ -255,6 +255,15 @@ def step(name, **kv):
         gc.temperature = 1.0
         gc.top_k = 0
         gc.top_p = 1.0
+        # STATE the penalty rather than inherit it. repetition_penalty is
+        # absent from generation_config.json — infer.py and api.py pass 1.1 at
+        # CALL time, so it belongs to the synthesis recipe, not the model. This
+        # dump does not pass it, so the codes here are penalty-free and the C++
+        # greedy path matches that deliberately (GenOptions::repetition_penalty).
+        # Pinning it means the two arms agree because both declare 1.0, not
+        # because both happen to inherit the same library default — which is
+        # the kind of agreement that quietly ends when a default changes.
+        gc.repetition_penalty = 1.0
 model.generation_config.max_new_tokens = MAX_FRAMES
 step("model_loaded", greedy=GREEDY, max_frames=MAX_FRAMES,
      dtype=str(next(model.parameters()).dtype))