docs: mutation-class catalog, padding-convention note, integer-only requant open problem, SYM pool mechanics

[LeoBuron]

Problem

Four pieces of decided-but-scattered knowledge have no home in docs/CONVENTIONS.md:

1. Mutation-class catalog — scale-blindness, eps-vacuity, layout-divergence, uniform-dy-vacuity, freeze-the-scale. Currently spread across generator docstrings, C test comments, and commit messages; every new SYM gold suite re-derives them from scratch.
2. SAME-vs-EXPLICIT padding note — C SAME is TF-minimal padding and diverges from PyTorch's symmetric padding for stride > 1 / even kernels; PyTorch-parity weight loading needs EXPLICIT. Bit us in the v2 examples; never written down.
3. Integer-only requant open-problem entry — Deutel et al. (arXiv:2407.10734) leave the requant realization open and use the FPU extensively; ODT's requant primitives are float-path. Per the research-backed-numerics rule an int-multiplier+shift scheme must NOT be homegrown — document it as an open problem with the literature pointer.
4. SYM pool mechanics note — Max = select, Avg = scale-fold, Adaptive = rounded integer division (layer: SYM_INT32 arms for MaxPool1d / AvgPool1d / AdaptiveAvgPool1d #205), in the style of the existing "two accumulation schemes" note.

Proposed work

One docs PR against docs/CONVENTIONS.md, parallelizable with everything; item 4 may land with the pools PR instead if M4 ships first.

Relations

#192, #137, #205.

🤖 Generated with Claude Code