0035 — A null pre-split drives Apply as Removed with invert, not Isolated

Status: accepted · Date: 2026-08-15

Context

BpePreTokenizer runs two patterns: a pre-split first, and a second pattern that re-splits the pieces it produced (#143). Only the pre-split carries a declared behaviour. A null pre-split still needs one, because Apply is driven by a behaviour whatever the caller declared, and the choice was made in a comment rather than recorded — #187 is what moved it here.

The two candidates look interchangeable, and are, whenever the pattern leaves no gap between its matches. Every shipped byte-level pattern — Gpt2, Llama3, Qwen2 — is of that kind.

Decision

The default is Removed with invert on — "keep the regex matches, drop everything else" — never Isolated.

BpePatterns.Whitespace is the pattern that separates them. Its \w+|[^\w\s]+ never matches a run of whitespace, so under Isolated that whitespace would surface as a piece of its own and reach the merge loop as an uncovered symbol. Measured against bpe.json — the " leading space" case — the reference produces no such piece, and no substituted token for one. Removed with invert keeps that path byte-for-byte unchanged.

Consequences

• The default is not exercised by BpeSplitBehaviorTests: every corpus case supplies its own BpeSplitStep, so the branch is reached only by a caller who declares no pre-split. A test that covers it would have to be written against the reference, not against the corpus.
• BpePreTokenizer carries a two-line comment pointing here, rather than the fifteen-line argument it used to carry inline.