test(dpt): use reproducible native-scale hash references

AGENTS.md

@@ -105,7 +105,7 @@ output.
   library layers, uses the reference model's `state_dict` names, takes NumPy arrays or
   `ServerArray`s as weights, and is not loaded by `import tensor_ee`. A random-weight model's
   numerical reference uses integer-generated weights with calibrated BatchNorm (CONTRIBUTING,
-  "Model references follow the same rule"). `compile_model` stays the
+  "Add a block", the paragraph "Model references follow the same rule"). `compile_model` stays the
   CNN path (`torch.fx` onto the R blocks): attention models go through the converter, which is
   extended with recognizers, never by routing them through `compile_model`.
 - **A new recognizer** follows CONTRIBUTING's "Add a recognizer":

CONTRIBUTING.md

@@ -107,8 +107,11 @@ model. [`tests/test_graph_identity.py`](tests/test_graph_identity.py) refuses to
 generators while the G1 fixtures build.
 
 **Model references follow the same rule.** A random-weight model's numerical reference,
-the PyTorch model its mirror is compared with, takes its weights from integer arithmetic
-(`_fixed_weights`), never from a library's native random initialization. Native
+the PyTorch model its mirror is compared with, takes its weights from integer arithmetic,
+never from a library's native random initialization: a counter-based integer hash of each
+parameter's name and position, identical on every platform, scaled per tensor to the
+standard deviation the library's own initializer gives it, so the reference behaves like a
+natively initialized model. Native
 initialization is not bit-identical across platforms, and it can let activations grow
 far beyond a trained model's. Its BatchNorm running statistics are then calibrated from
 a fixed set of integer-generated inputs, separate from the evaluation input: a fixed

docs/api.md

@@ -261,10 +261,13 @@ The caller prepares the crop and handles georeferenced relabeling.
 The [evidence summary](converter.md#smp-dpt-vit-b16-evidence-and-scope) keeps graph
 identity, offline mirror agreement, live agreement and model accuracy separate.
 Random-weight full-width and reduced conversions are byte-identical to their
-hand-composed targets. Full width is offline only: the DPT offline NumPy mirror
-misses the 1e-4 absolute target against both float32 and float64 PyTorch, with
-0 argmax flips. Float32 PyTorch itself differs from float64 PyTorch by more
-([numbers and dtype record](vit-b16-dpt224.md#numerical-scope)). Live agreement
+hand-composed targets. Full width is offline only: with counter-based integer-hash
+weights at native initializer scales and fixed BatchNorm calibration, head/logits
+meet 1e-4 against float32 PyTorch (full-width errors 2.06e-5/2.01e-5, reduced
+7.60e-6/7.29e-6; zero argmax flips). Every intermediate's error against float64
+PyTorch is no larger than float32 PyTorch's own; intermediate agreement within
+1e-4 against float32 is not claimed
+([scales, results and historical records](vit-b16-dpt224.md#numerical-scope)). Live agreement
 covers only the reduced encoder twin with rank-at-most-4 projection weights,
 the structured reduced DPT twin (`vit_dpt_reduced`, 7 requests, sampled head/logit
 readouts), and individual blocks' recorded settings. The full-width live run

docs/converter.md

@@ -613,20 +613,26 @@ one class token and cat readout, as in the example above. Four claims stay separ
   and [SMP DPT converter gate](../tests/test_smp_dpt_converter.py) convert random-weight
   models to graphs byte-identical to the hand-composed targets, at full width and
   reduced width. Both builds bind the same symbolic table reads; reduced inline
-  gates also compare random-weight literals. These gates establish graph identity,
+  gates also compare native SMP/timm random-weight literals, independently of
+  the calibrated integer numerical references. These gates establish graph identity,
   not numerical execution.
 - **Offline mirror agreement.** The [encoder evidence](vit-b16-encoder224.md)
   and [DPT numerical evidence](vit-b16-dpt224.md#numerical-scope) compare NumPy
   mirrors with PyTorch. The [reduced DPT fixture record](evidence/dpt-reduced-offline.json)
-  has largest absolute error 1.86e-9. Full width is **offline only** and does
-  **not meet the unchanged 1e-4 absolute target**: against float32 PyTorch,
-  fusion3 differs by 6.03e-4 and logits by 3.51e-4; against float64 PyTorch,
-  fusion3 differs by 1.18e-4 and logits by 1.50e-4. Activations reach about
-  190, where one float32 step is about 1.5e-5. Float32 PyTorch itself differs
-  from float64 PyTorch by more (fusion3 6.13e-4, logits 3.70e-4); there are
-  **0 argmax flips**. The [dtype record](evidence/dpt-dtype-offline.json)
-  retains these failures; neither rounding scale nor unchanged decisions relaxes
-  the absolute target.
+  has largest absolute error 1.86e-9, using its unchanged compact fixture weights.
+  The final [hash-reference record](evidence/dpt-hash-offline.json) uses a
+  counter-based integer hash at each native initializer's standard deviation,
+  followed by fixed four-batch BatchNorm calibration. Head/logits meet 1e-4
+  against float32 PyTorch: full-width errors 2.06e-5/2.01e-5, reduced errors
+  7.60e-6/7.29e-6. Every intermediate's mirror/float64 error is no larger
+  than float32 PyTorch's own. Full-width fusion3 peaks at 51.93 (float32
+  step 3.81e-6): mirror/float32 error 1.88e-4, mirror/float64 4.64e-5,
+  float32/float64 1.87e-4. This is an intermediate ordering claim, not
+  1e-4 intermediate agreement against float32. Every recorded output has zero
+  argmax flips. The [native-init](evidence/dpt-dtype-offline.json),
+  [unscaled periodic](evidence/dpt-integer-offline.json), and
+  [fan-in periodic](evidence/dpt-fan-in-offline.json) records remain unchanged
+  as superseded history. Full-width agreement remains offline only.
 - **Live agreement.** It exists only for the
   [reduced ViT encoder twin](evidence/live-fixtures/20261103T022505Z-vit_encoder_reduced.json)
   with **rank-at-most-4 projection weights**, the