feat: support DeepSeek V4 Flash 0731

[raullenchai]

What changed

• adds a deepseek-v4-flash-0731-mxfp4 alias for the 0731 MXFP4 MLX checkpoint
• implements the checkpoint-specific prompt encoder and native DSML tool parser
• translates the official checkpoint quantization paths to Rapid vendored-model paths
• normalizes Hugging Face AddedToken metadata in the fallback tokenizer loader
• prevents incompatible KV caches from retrying failed disk serialization every decode token
• records the exact 156 GB checkpoint size, 192 GB memory floor, and parser coverage exemption

Why

This checkpoint intentionally ships a dedicated encoder and DSML wire format. Without explicit support, Rapid falls back to incompatible ChatML, misloads mixed MXFP4/MXFP8 tensor shapes, and cannot return structured tool calls.

Validation

• official DeepSeek encoder fixtures 1 and 2 match byte-for-byte
• focused DSV4/parser/checkpoint suite: 42 passed
• parser coverage + streaming parity suite: 25 passed
• complete local suite: 14,904 passed, 174 skipped; 13 environment/server-required failures and one parser coverage failure discovered, with the parser coverage failure fixed and revalidated
• Ruff, type-check, diff check, wheel build, Python 3.10/3.11/3.12 CI, Apple Silicon CI, and PR validation
• final wheel force-installed and run outside the repository
• real 34-shard, 156 GB checkpoint loaded on a 256 GB Mac Studio
• basic inference: HTTP 200, correct Chinese answer (法国的首都是巴黎)
• long-prompt inference: HTTP 200 with 729 prompt tokens; failed KV serialization no longer hot-loops
• public rapid-mlx share health check: ready=true, model_loaded=true

Known limitation

The community MXFP4 checkpoint answers normal chat correctly, but tool-use quality was unstable in local E2E trials: chat mode declined a requested call and thinking mode repeated thinking to the token limit. The official DSML encoding and Rapid parser are covered byte-for-byte/unit/streaming, but this PR does not claim that this particular quantization preserves reliable tool-selection quality.

[github-actions]

PR #1364 validation scorecard

Title: feat: support DeepSeek V4 Flash 0731
Author: raullenchai
Diff: 14 file(s), +659/-6 LOC, blast radius: high

Verdict: MERGE-SAFE

step	status	summary	time
fetch	PASS	14 files, +659/-6 LOC, blast=high	1.3s
test_plan_check	PASS	no checkbox-style test plan found	0.0s
cl_description_quality	PASS	title OK + body has rationale (2036 chars)	0.0s
supply_chain	PASS	no hooks touched, no suspicious patterns, deps clean	0.0s
test_env_check	PASS	installed trusted-pins (2 missing plugin(s) recovered)	4.2s
codex_review	skip	codex CLI not found on PATH (install: npm i -g @openai/codex)	0.0s
lint	PASS	clean (12 file(s))	0.0s
targeted_tests	PASS	exit 2 (in 18 target file(s))	5.1s
diff_coverage	skip	diff-cover not installed — advisory coverage unavailable	0.0s