Struct Glm5SpecSession 

pub struct Glm5SpecSession {
    pub committed: Vec<u32>,
    pub rounds: usize,
    /* private fields */
}

glm5_next SERVED speculative session (lane/glm5-spec-routing, 2026-08-30): the state one request’s spec decoding carries across worker bursts — the dspark/gemma session twins’ shape. The session OWNS its trunk cache (the worker’s s.cache stays None); at every burst boundary the invariant is cache.pos == committed.len() with each committed row’s trunk state exactly what a plain prime of that sequence would hold (the accept walk’s basis, pinned by the tparallel gate), plus ONE emitted-but-uncommitted anchor token (the next round’s verify row 0 — the dspark last convention).

Fields

committed: Vec<u32>

Every token whose trunk state the cache holds, in order (prompt + committed generation). EXCLUDES the live anchor.

rounds: usize

Verify rounds completed over the session lifetime (the worker’s per-burst rounds-delta receipt, the dspark rounds convention).

Implementations

impl Glm5SpecSession

pub fn cache_max_ctx(&self) -> usize

Context capacity of the session’s cache (the server’s ContextFull guard).

pub fn pos(&self) -> usize

Trunk rows currently committed (== committed.len() at burst boundaries).

pub fn finished(&self) -> bool

EOS committed or the context guard tripped: the next burst would emit nothing.

pub fn demote_eligible(&self) -> bool

True when the session is a legal demotion source (loop-port fold-in, map #8): GREEDY only — a sampled session’s committed stream depends on its session-owned Philox counters, and the plain batched sampler is a different random program mid-request (the exact exclusion the MTP and dspark sweeps carry).