Module fingerprint 

Deterministic content hashing: token hashes, rolling k-grams, winnowing.

Every hash here is a pure function of token content (kind tags and normalized or raw text). No process randomness, no position, no file identity enters any hash, so runs are reproducible and equal content always collides intentionally. The 64-bit FNV hashes are used only for candidate indexing. Grouping uses a domain-separated 128-bit BLAKE3 digest, so an attacker-controlled FNV collision cannot combine unrelated findings.

Functions

kgram_hashes

Rolling polynomial hashes of every k-gram of units (mod 2^64): gram(i) = units[i]·B^(k-1) + … + units[i+k-1].

norm_sequence_hash

Content key of a normalized token sequence.

norm_token_hash

Hash one normalized token: kind tag, atom discriminant, atom payload.

raw_sequence_hash

Content key of a raw token sequence: the fold of its per-token hashes.

raw_token_hash

Hash one raw token: kind tag plus raw text.

winnow

Winnowing: select fingerprints from k-gram hashes.