nx_cdc_lib.nx

nx_cdc_lib.nx -- CONTENT-DEFINED CHUNKING (FastCDC gear, normalized) as a LIBRARY (/compare/dataio DI5, 2026-09-05). WHY: fixed-size chunking re-keys every chunk after an insertion, so a one-byte edit near the head of a 400 KB artifact makes every following chunk "new" to a receiver that already holds the previous version -- the whole file ships again. Content-defined boundaries (a rolling gear hash, cut where its low bits are zero) re-align within one chunk of the edit, so only the touched region is new. This is the FastCDC shape (Xia et al., USENIX ATC 2016, the dataio board's pinned reference): a 256-entry gear table, a one-shift-one-add rolling fingerprint, a minimum length skipped outright, NORMALIZED chunking (a stricter mask before the average, a looser one after, so lengths cluster around the average) and a hard maximum. EVERYTHING IS DERIVED, NOTHING TUNED. The gear table is SHA-256 of a NAMED seed and the byte value -- reproducible on any host, so no table ships that could drift. The mask bit counts come from log2(avg). min, avg and max belong to the CALLER: nx_content_put derives them from the ONE wire chunk constant it already owns (max = the wire chunk, so every content-defined chunk still fits one call). HONESTY LIMITS, stated: the masks take the LOW bits of the fingerprint (the paper spreads the bits; the low-bit variant is simpler and its statistics are PROVEN by nx_cdc_gate on a pseudo-random corpus rather than assumed). The final chunk may be shorter than min. This lib never reads a file or the clock; it computes over a buffer. license_tier: ORIGINAL No hw writes (Rule 26).