Add ECP-7 Batch 27 collision pulse study

[FatosLocos]

Summary

• preregister and run one late bounded direct sender-collision pulse on the stable Batch 25 base
• preserve the complete Batch 25 trajectory through step 30,000 and reuse previously calibrated collision-loss settings
• record the sealed paired seed-11 result, diagnostics, and compact evidence manifest
• close coefficient and timing development for the existing relaxed collision-pair loss
• update fresh-agent guidance toward an identity-preserving sender-representation experiment

Result

ECP7-B27-I reached 85.38% mean train and validation exactness and 85.35% universal-translator validation exactness. The collision-pair banks became smaller and message entropy increased, but the sender codebooks retained exactly 13,440, 12,960, 12,720, and 13,440 unique messages. Worst-link validation regressed to 82.13%, so both the unchanged development gate and the added direct-capacity criterion failed. The ECP-7 confirmatory test remained sealed.

Verification

• .venv/bin/pytest -q (127 passed)
• .venv/bin/ecp7 --config config/ecp7-b27-development.yaml validate
• paired seed-11 smoke, develop, and 100-permutation analyze runs
• 151 shared trajectory records through step 30,000 match Batch 25
• all tracked Batch 27 manifest hashes verified against local artifacts