release: v0.2.2 — MLX-parity quality + structured output + LoRA lifecycle

[ohdearquant]

Summary

Release notes for v0.2.2 — 58 commits, 29,524 insertions, 1,012 deletions since v0.2.1.

Workspace version is already at 0.2.2 (bumped during the release cycle), so this PR adds only the release-notes file.

Major themes

1. MLX-parity quality on Qwen3.5-0.8B: PPL gap 0.77 → 0.029 (RoPE stride-half fix fix(inference): RoPE stride-half pairing — WikiText-2 PPL gap 0.77 → 0.029 vs MLX #96, FP16 lm_head, asymmetric Q4)
2. Structured output: XGrammar engine (ADR-046)
3. Serving: continuous batching with chunked prefill (ADR-048), prefix-shared paged KV cache (ADR-047)
4. Multimodal: vision encoder (ADR-049)
5. MoE: Metal dispatch with expert coalescing (ADR-053)
6. Speculative decoding: GDN self-spec draft, MTP on QuaRot Q4, probabilistic rejection sampling (ADR-050)
7. Fine-tuning: LoRA full-lifecycle (ADR-057 D1-D5, feat: implement ADR-057 LoRA full-lifecycle consumer API (D1-D5) #65), Adam/AdamW + LoRA gradients (feat(tune): Adam/AdamW optimizer + LoRA gradient computation #90)
8. Performance: CPU/embed SIMD (perf(embed): SIMD optimizations + correctness hardening #79, perf(inference): CPU SIMD optimizations + parity regression tests #80), Metal zero-copy decode + MLP fusion (fix(inference): Metal GPU build repair + zero-copy decode + MLP fusion #81), CI perf-regression tracking (ADR-058, feat(ci): ADR-058 CPU perf regression tracking + bench tooling #83)

Post-merge steps

git checkout main && git pull
git tag -a v0.2.2 -m "v0.2.2"
git push origin v0.2.2
gh release create v0.2.2 --title "v0.2.2 — MLX-parity quality" --notes-file docs/releases/v0.2.2.md
# Optional: make publish  (crates.io)

Test plan

• cargo check --workspace --features "f16 metal-gpu" clean
• Release notes match commit history (verified via git log v0.2.1..HEAD)

🤖 Generated with Claude Code