The matcher: one left-to-right walk producing slug text plus the spans it came from.

The algorithm, in one place

Everything positional is counted in grapheme clusters.

Pre-pass (once):

1. locale-aware lowercase (:turkic / :greek / default)
2. normalise to NFC — every later step reads NFC text, nothing re-normalises
3. compute word boundaries by Unicode property, before any mapping

Main pass — at each position, gather every rule from L2, L1 and L0 whose key matches here, then:

longest key wins
tie on length      -> higher layer wins (L2 > L1 > L0)
tie within a layer -> first declared wins
nothing matched    -> retry against the grapheme's canonical base
still nothing      -> strip it

Why longest-match and layer precedence cannot fight

They look like competing rules. They cannot disagree, by construction:

• L0 keys are always one grapheme — enforced by the generator.
• L1 and L2 never both match — L2 is Latin-only, L1 keys come from non-Latin script blocks.

So a key longer than one grapheme only ever comes from one layer, and cross-layer competition is always a length tie, settled by precedence. German ö→oe (L2) cannot be beaten by a longer L0 rule, because no such rule can exist.

Base-retry, and why it must come after the precomposed lookup

A grapheme with no rule is retried against its canonical base (NFD, marks dropped) before being stripped. That is what lets the Greek table omit ά έ ή ί ό ύ ώ — they decompose to base vowels that are keyed.

The ordering is load-bearing. Decompose first and German ships o instead of oe, because NFD turns ö into o + combining diaeresis and the de table keys the precomposed U+00F6. Retry is the fallback, never the first move.

Spans

Each emitted chunk records the rule output that produced it, so truncation can refuse to sever щ→shch into shc. Spans are carried through separator collapsing and trimming, because those steps move every offset.