v0.4.0

What's Changed

• feat(serve): hardening for concurrent requests under apc and speculation by @asher in #62
• fix(serve): reply-content indent loss and float16 sampling/softcap hardening by @asher in #63
• fix(serve): memory accounting and admission hardening; qualify mlx 0.32.1 by @asher in #64
• serve: manage memory pressure across admission, decode, and cache tiers by @asher in #67
• feat(dflash2): DFlash 2 drafter support for Qwen3.8-27B and Muse-Glimmer-30B by @asher in #65
• fix(serve): eviction headroom leak and cpu-stream bus error by @asher in #69
• fix(serve): evicted streamed models stayed resident by @asher in #70
• fix(thinking): mlx-vlm 0.6.15 thinking-off regression + end-to-end thinking control plumbing by @asher in #71
• fix(apc): lone-harvest patch matches the 0.6.15 harvest signature by @asher in #73
• fix(serve): governor kernel floor + APC retirement/kv-quant/ring-copy fixes by @asher in #74

Full Changelog: v0.3.2...v0.4.0