[initial-d]

Dense me_fasttext .bin models can now be converted into immutable native serving bundles and queried through the original FastText facade. The optional make fasttext-native build accepts either .bin or a bundle directory with the existing word/sentence/classification CLI commands. Conversion uses the actual registered DATrie IDs, original structural planner and mark-compact forwarding; layout/exact modes preserve all input bytes and original float32 accumulation order.

The original build targets and dictionary save/load overloads remain available. Native input uses bounded gather chunks and a decoded cache, while dictionary/output state stays resident. Training and quantization use the retained dense model. This is Linux x86-64 source compatibility for this repository's unquantized models, not an old C++ binary ABI or official fastText/.ftz/.z converter.

Loaded dictionaries now skip new OOV rows beyond the trained matrix, direct subword helpers use registered DATrie IDs instead of hash buckets, HS prediction rebuilds its output tree, and zero-vector string sentence queries avoid NaN. Equivalence to historical valid queries is tested separately from these corrected behaviors.

Validation:

• Actual legacy SGNS training on transformed existing enwik8: 90,380 × 64 float32 input rows, full reconstruction SHA equality, and identical held-out word/sentence bytes in nine fresh-process dense/layout/exact runs.
• Historical main (bea4302) vs new dense vs native: identical outputs on 1,024 registered-word and 128 sentence queries using the same actual trained model.
• Linux ASan/UBSan/LeakSanitizer facade fixtures: all input rows, UTF-8/OOV, softmax/HS scores, repeated/chunked float32 adds, immutable/failure/reload guards. Bundled trie archives remain uninstrumented.
• Installed wheel checks validate the converter import/entry point without Torch. Final exact-head CI includes the new sanitized facade job.

Measured limits: structural native input saves 3.18%, but the whole bundle saves only 0.46%. Exact sharing finds zero identical rows; matched identity row order is 103,371 bytes smaller. Native hot word/sentence queries are slower than dense, and first-pass decoding is substantially slower. Anonymous RSS falls about 18.9 MiB while total RSS falls only 1.34 MiB after touched mmap pages. This is functional fusion evidence, not a universal compression/speed claim or an Engram task-quality result.

Usage · Raw trained results, matched layout controls, historical comparison and preserved initial failure. All remote work used CPU only in the maintainer's directory; no GPU process was touched.