feat(early-exit): CALM-style adaptive per-token threshold + arch gating (v0.27.0)

[konjoinfinity]

Konjo sprint — v0.27.0

Fourth sprint from this session's Discovery sweep (#15 CyclicJudge, #16 reliability, #17 KV eviction — all merged). Continues the inference-optimizer half of kairu picked back up in #17.

Why

EarlyExitDecoder used a single static confidence threshold for every token — but an early-token mistake propagates through the whole sequence, so early and late tokens shouldn't share one bar. And AutoProfile recommended early_exit for any mid-size model regardless of whether the architecture can actually exit (encoder-only models have no left-to-right confidence trajectory; shallow models have nothing to skip).

What

kairu/early_exit.py — CALM-style adaptive threshold (Schuster et al. 2022), opt-in via adaptive=True (static path bit-for-bit unchanged):

• The bar decays geometrically over decoding steps:
  threshold(t) = min_confidence + (confidence_threshold − min_confidence)·exp(−adapt_decay·t).
  Early tokens must clear a high bar; it relaxes toward min_confidence as decoding proceeds. New effective_confidence(step) exposes the schedule.
• min_confidence is clamped to the base so the schedule is monotone non-increasing. The constructor now validates inputs (it previously accepted any value). Stats gain adaptive + final_confidence_threshold.

kairu/auto_profile.py — arch-suitability gating:

• New _early_exit_suitable(model, name) rules out encoder-style families (BERT/RoBERTa/ELECTRA/DeBERTa/T5/encoder) and sub-6-layer models. Both early_exit and layered_early_exit recommendations fall back to vanilla with an explanatory rationale when the gate fails. (Correctly resolves num_layers as a method on LayeredModelInterface.)

kairu/wrapper.py — adaptive_early_exit: bool = False flag (parallel to adaptive_gamma) forwards to the decoder, making the feature reachable from wrap_model.

Quality gates

• 732 passed, 4 HF-gated skipped (+17 new).
• New tests/test_early_exit.py takes the module 21% → 100% coverage — validation, the decay schedule, every exit reason (confidence / entropy / max_tokens), and an adaptive-vs-static divergence case with hand-derived expected values (threshold(2)=0.5+0.4·e⁻¹≈0.647). 3 arch-gate tests added.
• early_exit.py / auto_profile.py: ruff lint + format clean; recommend stays grade C cc=15 (the documented ≤15 ceiling, same grade as before this change and consistent with existing _collect_hardware/paired_t_test).

Reuse note

Pure NumPy + stdlib — no HF/torch in the decode path. The adaptive schedule generalises the existing static threshold (which is just the adaptive=False branch); the arch gate is extracted as a single helper so recommend stays one readable decision tree.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TmSqgYqcYfzPnkpDVR5vdw

---

Generated by Claude Code