[initial-d]

Large n-gram tables can share training updates while their frozen vectors remain different. Add a CPU workflow that audits actual keyed rows, then measures the original DATrie/mark-compact contribution against raw, native identity, lexical layout and exact encoded-dedup controls at the same encoding.

On the unchanged 1,855,907 × 256 finite SGNS table, conditional identical-word-support updates are almost collinear, but final-vector cosine is only 0.21–0.35. Complete FP32/FP16 structural savings are 0.12–0.63%; INT8 remains larger than raw INT8, and broad native query misses remain slower. Quantization/joint sharing are explicitly lossy; all encoded representative bytes and original IDs are independently verified. This is storage/vector evidence, not LLM task-quality or convergence evidence.

• 27 fully exported layouts, every original ID verified, and 108 fresh query workers across two randomized rounds per trace. No GPU use or retraining.
• Actual initialization and original training-ID provenance explain the co-training signal; no synthetic vectors are used for compression evidence.
• Complete raw reports, four figures, source/environment fingerprints, retained launcher failure and partial-controller handoff, checks and reproduction commands: evidence report.
• Validation: 16 current targeted tests pass; Linux full-suite snapshot has 145 tests/17 skips, local snapshot 145/26 skips. Final-head CI at e5950af755238c7e0706089d1112e4cd3d699afb: 18/18 passed before merge. Fabricated correctness fixtures remain separate from measured data.