Record ECP-7 Batch 17 global collision replay

[FatosLocos]

Summary

• add deterministic training-only global collision mining and replay
• preserve the Batch 15 trajectory through step 15,000 before activating replay
• preregister and run ECP-7 Batch 17 on the sealed development split
• record the valid negative result with verified artifact hashes
• keep the confirmatory test sealed and recommend one attenuated-coefficient follow-up

Result

Batch 17 reaches 83.15% train exactness, 77.09% validation exactness, and 83.01% universal-translator validation exactness. Global collision replay modestly increases unique-message use and reduces collision meanings, but full replay weight overpowers the remaining task objective, lowers sender agreement, and causes collision churn. No sender becomes injective, and Batch 15 remains the strongest weak-structure base.

Verification

• 100 passed
• active replay-path integration check passed
• ecp7 validate: 14336/1024/1024
• both analyzers passed with 65/65 artifact hashes
• 16,384 validation-only episodes per run
• no confirmatory-test metrics or episodes
• valid exact 14-bit local channel
• manifest JSON and artifact hashes independently verified
• git diff --check