[initial-d]

Existing FP32 structural ablations do not establish whether DATrie prefix/suffix sharing adds value after ordinary weight quantization. This adds frozen CPU FP16 and scale-aware per-row INT8 storage, preserving original IDs and complete encoded representative bytes through the native mark-compact path.

The same 18 cached trained SGNS/GRU tables now have 486 matched storage variants: raw mmap, native identity/layout, exact encoded-byte dedup, and structural sharing at five cosine thresholds. Results count scales, maps, metadata and actual key artifacts; whole GRU bundle costs and source hashes are retained. Selection is locked on validation before testing.

At the fixed 0.99 cosine/norm guard, INT8 still leaves an additional 20–28% saving on SGNS and 11%/15% on order-8/12 GRU versus the best same-encoding control. Order-4 GRU INT8 compaction increases storage. These are approximate row changes on small public-text trained references, not official trained Engram or general LLM evidence; numeric gathers remain slower than raw encoded mmap.

Seven of 27 primary validation choices exceed the held-out KL budget and remain failures. A separately labeled post-main-run fixed-0.99 audit passes all 27 tests without replacing the primary choices. Rate–distortion plots, per-seed data, commands, environment, build/test logs and the initial launcher failure are inspectable in docs/quantized_structural_evidence_20261031/.

Validation: 137 local unit tests pass (26 platform/optional skips); 24 pinned Linux codec/DATrie/causal tests pass; every encoded row and original-ID forwarding result is verified in the main run. The subsequent max-float32 INT8 overflow guard passes extreme-value checks and preserves all 54 trained encodings byte-for-byte.