Add native Qwen3.8 27B model lanes

[a-github-name]

Summary

• add vision-chat-q38-27b, pinned to Qwen's official BF16 Qwen3.8-27B snapshot
• add vision-chat-q38-27b-4bit, combining LM Studio's pinned MLX 4-bit target with the official pinned BF16 MTP shard mounted under mtp/
• extend the native Qwen-family runtime for Qwen3.8 dense text, code, vision, tokenizer/template, published sampling, context, and image-sizing behavior
• keep dense MTP greedy speculation explicitly opt-in and disabled for sampled or JSON-constrained generation
• record composite source provenance in managed manifests and cover catalog validation, memory admission, CLI parsing, coding benchmarks, and docs

Validation

• ./scripts/check.sh
  • strict SwiftLint: 1,237 files, 0 violations
  • XCTest: 2,859 executed, 207 checkpoint-gated skips, 0 failures
  • Swift Testing: 30 passed
• rebased focused suites: 308 executed, 3 skips, 0 failures
• release build passed
• managed pull preflight recognized both pinned component sources and reported the 4-bit composite ready to pull
• canonical managed-layout smoke reported a valid 19.47 GB composite manifest with primary and component provenance

Real-checkpoint evidence

Measured on an M4 Max coding smoke:

Lane	Target-only warm decode	MTP warm decode	Speedup	MTP acceptance
BF16	8.78 tok/s	18.81 tok/s	2.14x	75.6%
4-bit	26.5 tok/s	38.8 tok/s	1.46x	75.6%

A deterministic 24-task stride through the official HumanEval dataset scored 20/24 for both 4-bit target-only and 4-bit + MTP, with the same four failures. One failing case produced 177 target-only tokens versus 174 with MTP, so the fast path remains opt-in rather than claiming exact greedy-output parity.

Operational notes

• BF16 pull size: 55.59 GB; 64 GB minimum unified memory, 96 GB recommended
• 4-bit composite pull size: 19.47 GB; 32 GB minimum unified memory, 48 GB recommended
• enable experimental dense MTP with MERERUN_Q35_MTP_SPECULATION=1
• default, sampled, and constrained requests retain target-only decode