Stage 2 decisions - Superfloat Python

Question	Choice here	Where
Default scale policy (spec section 12 open question)	weights per tensor by default, per channel on request (`scale=`); activations dynamic per call; float scales on the host	`torch/_core.py`
Headroom	`"bound"` (default): raise the activation scale until the L1 bound proves no partial sum or bias add can saturate; `"none"`: raw scale, saturation counted. Measured: without headroom a trained MLP drops from 0.95 to 0.18 accuracy	`torch/_core._activation`
Matmul bound	tightened from the spec’s `K max\	A\	max\	B\	`to`min(max_i \	\	a_i\	\	_1 max\	b\	, max\	a\	max_j \	\	b_j\	\	_1)`, never looser, exact integer arithmetic; used by both pre-flight and the headroom search, so what the host proves safe the device pre-flight proves safe too	`_linalg.matmul_bound`
Bias	added after the chain as one saturating add (`addmm` semantics of 1.1), quantized in the accumulator scale `a s_w`; a bias that does not fit is counted as `bias_saturated`	`torch/_core.linear`
Gradients	straight-through: `quantize` passes the gradient where the input was in range, 0 where it saturated; `linear` / `matmul` use the float gradient with the dequantized weight	`torch/_ops.py`
Output dtype	float32 for float16 / bfloat16 / float32 inputs, float64 for float64 (bfloat16 and float16 cannot hold 32767/32768)	`torch/_ops.out_dtype`
Conv2d	unfold on the host (data movement), then the SF linear chain; groups != 1 and string / non-zero padding are skipped and reported	`torch/_modules.SFConv2d`
Modules not converted	Conv1d/3d, transposed convs, MultiheadAttention, RNNs, Bilinear: skipped with a reason; custom attention uses `sft.matmul`	`torch/_modules.convert`
`last_status`	process-wide registry keyed by layer name; `last_status(model)` filters by that model’s layers	`torch/_status.py`
Compile backend	retargets `sf::linear` / `sf::matmul` nodes to a `DeviceRunner` on `sf.config(default_device)`; unsupported widths stay host-emulated; inference-only (raises with grad enabled)	`torch/_backend.py`
`autocast` as a decorator	sync functions only; torch function modes are thread-local, so a coroutine or generator would leak the mode across awaits	`torch/_autocast.py`
NaN / ±inf in SF layers	routed through the SF policy before any scale is chosen, like a Stage 1 cast: NaN is `invalid` (device value 0), ±inf is `overflow`; under a non-raising policy ±inf stands in as the largest finite magnitude of the tensor with its sign, so it saturates in the scale the finite values choose	`torch/_core.sanitize`
Conversion and parameters	`SFLinear.from_float` / `SFConv2d` wrap the source layer’s own `Parameter` objects (identity, `requires_grad` and training state preserved), so `convert(..., inplace=True)` keeps a pre-existing optimizer connected; a non-inplace `convert` deep-copies first	`torch/_modules.py`
Shared / aliased modules	converted once per distinct module; every alias (including inside nested containers) is replaced by the same SF object; reported under `shared`	`torch/_modules.convert`
Root `Conv2d`	converted like a root `Linear`; an unconvertible one is reported as skipped	`torch/_modules.convert`
Huge float64 values	the bias scale is derived with overflow guarded (`a_bias` from finite ratios, a bias no finite scale can encode is counted as `bias_saturated`); a result outside float64 is `inf`, as in float	`torch/_core.linear`
Request-scoped status	`sft.capture_status()` collects statuses for the current thread or task (context-local), alongside the process-wide `last_status()`; use it for concurrent serving where models share layer names	`torch/_status.capture_status`
Research `sf_quantize`	the spec says to promote it from the superfloat research code; that code is not in this repository, so `quantize` is written from the Stage 1 cast and gated on producing the same codes. Replace or cross-check when the research code is available	`torch/_core.fake_quantize`