Summary

• ASI coherence tracking (crates/zeph-llm/src/router/asi.rs new): per-provider sliding window of response embeddings; coherence score (cosine similarity vs window mean) penalizes Thompson beta and EMA score when responses drift; emits tracing::warn! on ASI drift below configurable threshold
• Quality-gate cascading (router/mod.rs): after Thompson/EMA selection, optionally compute cosine similarity between query and response embeddings; below threshold escalates to next provider in pool; on exhaustion returns best response seen; fail-open on embed errors
• Bounds validation for quality_gate (must be in (0.0, 1.0]) and ASI coherence_threshold/penalty_weight (clamped with warn log)
• 4 new unit tests covering gate happy path, exhaustion fallback, and guard logic

Config

[llm.routing]
quality_gate = 0.75   # optional, disabled by default

[llm.routing.asi]
enabled = false
window = 5
coherence_threshold = 0.7
penalty_weight = 0.3

Test plan

• cargo nextest run --workspace --lib --bins — 7690/7690 pass
• cargo clippy --workspace -- -D warnings — clean
• cargo +nightly fmt --check — clean
• Live session test with quality_gate enabled — verify escalation logged at info
• Live session test with ASI enabled and window=3 — verify drift detection fires

Closes #1841, #2165