chore(release): 0.2.4

[Pushkinist]

Release 0.2.4 — version bump + CHANGELOG.

Vision/KV/embedding-lookup bug-fix batch for Qwen3-VL and Gemma 4, plus a /metrics/cache fix and the Homebrew bottle build+publish flow. Bundles the six fixes merged since v0.2.3:

• /metrics/cache: response shape diverges from docs (.models[] not .prompt_cache; itl not last_itl); .ttft ring empty for non-streaming #141 /metrics/cache TTFT for non-streaming + docs (fix(server): record TTFT for non-streaming completions + align /metrics/cache docs (#141) #142)
• Homebrew bottle: binary install + clean mlx-c dependency handling #139 Homebrew bottle build+publish flow (feat(release): Homebrew bottle build+publish flow (#139) #143)
• gemma4 e4b QAT (mxfp4/nvfp4) image inference degenerate — vision conditioning dropped (text + mxfp8 vision fine; 26b QAT fine) #140 Gemma 4 image grounding — tokens inside the user turn (fix(server): place image tokens inside the user turn (gemma4 vision grounding) (#140) #144)
• Qwen3-VL image inference ignores --max-ctx: KV cache stays at 4096, large images fail with slice_update broadcast #138 Qwen3-VL KV sized from --max-ctx (fix(qwen3-vl): size KV from --max-ctx + chunk prefill (#138) #145)
• Qwen3-VL large-image ViT full-attention overruns Metal GPU watchdog (>~1300 soft tokens) #146 Qwen3-VL O(seq) on-device embed lookup + ViT tiling (fix(qwen3-vl): O(seq) on-device embedding lookup + ViT attention tiling — large images no longer hit the Metal watchdog (#146) #147)
• Qwen3.6 (qwen3_5_moe) embed_lookup uses O(seq²) eye(seq)@w on CPU — port the #146 on-device take+dequantize fix #148 qwen3_5_moe on-device embed lookup (fix(qwen3_5_moe): on-device take+dequantize embed_lookup (drop O(seq²) eye@w CPU trick) (#148) #149)

No breaking changes. make ci green locally (fmt + clippy + test + deny + audit).

🤖 Generated with Claude Code