Trait SpecConstraint 

pub trait SpecConstraint {
    // Required methods
    fn mask_logits(&mut self, logits: &mut [f32]) -> Result<(), String>;
    fn mask_words(&mut self) -> Result<Vec<u32>, String>;
    fn is_allowed(&mut self, tok: u32) -> Result<bool, String>;
    fn consume(&mut self, tok: u32) -> Result<(), String>;

    // Provided methods
    fn draft_mask_enabled(&self) -> bool { ... }
    fn draft_begin(&mut self) -> Result<(), String> { ... }
    fn draft_mask_words(&mut self) -> Result<Option<Vec<u32>>, String> { ... }
    fn draft_advance(&mut self, _tok: u32) -> Result<bool, String> { ... }
}

GRAMMAR HOOK for constrained spec decode (lane/constrained-full, 2026-08-03). The engine stays llguidance-agnostic: the server adapts its per-session grammar state behind this trait. CONTRACT (the verify-side truncation rule — token-identical to constrained plain greedy decode): the exactness walk runs UNMASKED first; the hook then (a) truncates acceptance at the first grammar-illegal accepted token, and (b) when the truncation fired or the bonus is illegal, the engine recomputes that slot as the MASKED argmax of the target’s own verify column (an unmasked argmax that is grammar-legal IS the masked argmax — masking only removes tokens — so the common case pays nothing). consume advances the state with each EMITTED token in order; EOS handling is the implementor’s job (skip).

Required Methods

fn mask_logits(&mut self, logits: &mut [f32]) -> Result<(), String>

-inf the current state’s banned ids on a HOST logits row (prompt-tail / init-feed masked argmax).

fn mask_words(&mut self) -> Result<Vec<u32>, String>

Packed 32-bit bitset words of the CURRENT state’s allowed set (device-mask form).

fn is_allowed(&mut self, tok: u32) -> Result<bool, String>

Is tok consumable in the CURRENT state?

fn consume(&mut self, tok: u32) -> Result<(), String>

Advance the state with an emitted token.

Provided Methods

fn draft_mask_enabled(&self) -> bool

Is draft-side masking available on this hook? Probed ONCE per burst, before the draft graph is captured (the mask is an in-graph node — its presence is a capture-time shape).

fn draft_begin(&mut self) -> Result<(), String>

Start a draft chain: clone the CURRENT (committed) grammar state into the speculative slot. Called once per spec round, before the first draft position.

fn draft_mask_words(&mut self) -> Result<Option<Vec<u32>>, String>

Packed 32-bit bitset words of the SPECULATIVE state’s allowed set (target-vocab ids), for the draft position about to be sampled. None = draft masking off (no-op).

fn draft_advance(&mut self, _tok: u32) -> Result<bool, String>

Advance the SPECULATIVE state with a PROPOSED draft token. false = the chain cannot continue (EOS proposed, or an unmasked position proposed something illegal) — the engine stops drafting; the token already pushed still goes through verify.

Dyn Compatibility

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors