Module layer 

Layer taxonomy for dictionary entries.

Every (code, word) belongs to exactly one Layer determined at build time. The layer feeds two purposes:

1. Coarse ordering — LAYER_BASE gives each layer a numeric base weight so 一级简码 always outranks any 二级简码, etc., regardless of per-entry frequency.
2. User-tunable preference — WubiDict::set_layer_pref lets the host multiply a layer’s contribution at lookup time without touching data.

Index values pack (layer << FREQ_BITS) | freq_score so the runtime can read both in one stream pass. freq_score is the corpus-derived frequency from data/weights/weights.tsv (capped at MAX_FREQ_SCORE; real data tops out around 50k), normalized within layer.

Enums

Layer

Discriminants are ascending priority: Auto = 0 is lowest, Jianma1 is highest. This makes the packed (layer << FREQ_BITS) | freq compare correctly with raw u64 ordering — higher u64 = higher priority — so the build-time merge step can keep the larger value on collision without special casing.

Constants

DEFAULT_LAYER_PREFS

Default layer_prefs, indexed by Layer as usize. Auto is dampened to 0.7 so extension characters don’t pollute the top of common 4-letter codes; everything else is 1.0.

LAYER_BASE

Per-layer base weight, indexed by Layer as usize (ascending). Values are spaced so that any in-layer frequency score (capped well below the gap) cannot reorder layers, but a sufficient layer_pref multiplier can.

LAYER_COUNT

Total number of layers. Acts as the array length for LAYER_BASE, DEFAULT_LAYER_PREFS, and any per-layer table the host might keep.

MAX_FREQ_SCORE

Largest freq_score the packed value can represent (2^FREQ_BITS - 1). The single source of truth for the freq domain — tests and any caller that needs to clamp/validate import this rather than hardcoding a copy (a stale copy in the proptest survived the E1 FREQ_BITS 56→20 change and silently broke the invariants until proptest caught it).

Functions

pack

Pack (layer, freq_score) into a single u64 index value. freq_score is saturated to MAX_FREQ_SCORE if it exceeds the field, never wrapped: for an order-by-value structure a wraparound would invert priority (a very high freq would pack to a tiny value and rank last), whereas clamping keeps “higher freq → higher-or-equal priority”. Real freqs (≤ ~50k) are far inside the field, so this only matters as a defensive guarantee.

unpack

Reverse of pack. Unknown layer bytes fall back to Layer::Auto (lowest priority) — preferable to panicking on a corrupt FST.