[noelEOS]

Base: 7eb3e47. Docs only: CONTRIBUTING.md and AGENTS.md.

What changes

• The 1e-4 target's scope. The 1e-4 absolute target against float32 PyTorch applies to every block fixture's outputs and to a model's final outputs (logits, and probabilities where produced). CONTRIBUTING already called it the "logit target"; this states the scope explicitly.
• Intermediate outputs of a model are recorded and checked against a float64 run of the same PyTorch model, with the same float32 weights and input. The mirror's largest absolute error against float64 must be no larger than float32 PyTorch's own.
• Every comparison records its scale:
  • the largest absolute value;
  • the float32 step at that value;
  • the errors against float32 and float64 PyTorch;
  • argmax flips.
• Random-weight model references use integer-generated weights, as fixtures already do, never a library's native random initialization. Their BatchNorm running statistics are calibrated from fixed integer-generated inputs, by a procedure fixed in advance and run once. Trained models keep their trained weights, and records made with native initialization stay as history.
• Claims: CONTRIBUTING now keeps four claims apart, as AGENTS.md already does.

Why

The full-width SMP DPT target, built with its native random initialization, reaches activations near 190 in its late decoder. There, one float32 step is about 1.5e-5. The offline mirror misses 1e-4 against float32 PyTorch (fusion3 6.03e-4), but float32 PyTorch is itself 6.13e-4 from float64 PyTorch, while the mirror is 1.18e-4 from it (docs/evidence/dpt-dtype-offline.json).

Offline diagnostics (not committed) showed that:

• with BatchNorm statistics calibrated from data, the logits come within 1e-4 of float32 PyTorch, and every output comes within 1e-4 of float64 PyTorch;
• the remaining misses are in intermediate fusion maps, where float32 PyTorch is as far from float64 as the mirror is from float32 PyTorch;
• native initialization builds different weights on macOS and Linux from the same seed.

The DPT reference will be moved to the new procedure in a follow-up.

Checks

• tests/test_docs.py, tests/test_history_citations.py, tests/test_x08_docs_and_api.py: 25 passed.
• No code, test, golden or evidence change, so G1, G1-ext and the catalogue are unaffected.