Add ECP-8 Batch 1 surplus-capacity study

[FatosLocos]

Summary

• preregister ECP-8 as a new sealed research phase with a fresh deterministic split
• compare a factorized 14-bit positive control, a 14-bit weak-structure paired control, and the identical weak system with a 16-bit global channel
• preserve the preregistration in a separate commit before all training results
• record Batch 1 as a valid negative development result without accessing the confirmatory split
• extend the post-hoc channel audit to validate global fixed-length messages

Result

The positive control remains perfect. Relative to the paired 14-bit control, the 16-bit intervention raises mean train exactness from 79.55% to 98.76%, worst-link validation from 64.84% to 73.34%, translator validation from 75.20% to 81.05%, and the minimum sender codebook from 12,048 to 15,095 messages. It still retains 225–265 collisions per sender and misses both registered 80% validation thresholds, so the development gate fails and the ECP-8 confirmatory split remains sealed.

Verification

• .venv/bin/pytest -q (138 passed)
• all three ECP-8 configs validate with split sizes 14336/1024/1024
• 100-permutation sealed post-hoc analysis for every arm
• 65 matching artifact hashes and 16,384 validation-only episodes per arm
• global channel audit checks every intervention message for four symbols in 0..15 and exactly 16 bits
• tracked Batch 1 manifest hashes verified against all local artifacts
• git diff --check