CI: gate MTP acceptance rate alongside gsm8k accuracy

[junhaha666]

What

Adds a MTP acceptance-rate gate to the native ATOM accuracy CI (atom-test.yaml), so a spec-decode regression fails the job instead of silently passing.

Why

The pipeline already scrapes MTP/spec-decode acceptance from the server log during the gsm8k run (atom_ci_metadata.mtp_acceptance_rate) and publishes it to the benchmark dashboard, but nothing gates on it. gsm8k accuracy alone cannot guard MTP: speculative decoding is lossless w.r.t. the target model, so a broken draft head leaves accuracy unchanged and only craters acceptance/throughput.

Concrete example this catches: DeepSeek-R1-0528-FP4 MTP was stable at ~61% from 6/16 through 7/8, then dropped to ~36% on 2026-07-08 and has stayed there — invisible to the accuracy gate.

Changes

• .github/workflows/atom-test.yaml — new Check MTP acceptance rate step after Check accuracy test results. Reuses the same gsm8k pass (no extra inference run), reads the recorded rate, fails if it drops below the per-model threshold or is missing (missing = MTP stats never emitted = treat as regression). Only runs for models that declare mtp_accept_threshold.
• .github/benchmark/models_accuracy.json — mtp_accept_threshold added to all 10 ATOM-backend spec-decode entries, set to each model's dashboard-observed minimum minus 1.5 points. FP4 MTP uses its healthy pre-regression minimum, so the current ~36% is caught rather than baked in.
• .github/benchmark/schema/accuracy_catalog.schema.json — adds mtp_accept_threshold / mtp_per_pos_threshold so validate_catalog.py accepts the new fields.

Thresholds (fraction 0-1, = hist min − 1.5 pts)

model	hist min	threshold
DeepSeek-V4-Pro MTP	63.70%	0.622
DeepSeek-R1-0528 MTP	63.50%	0.62
DeepSeek-R1-0528 MTP Online-Quant	59.74%	0.582
DeepSeek-R1-0528-FP4 MTP	60.84%*	0.593
Kimi-K2.5-MXFP4 Eagle3	68.52%	0.67
GLM-5.2-MXFP4 MTP	61.02%	0.595
Qwen3.5-397B-A17B-FP8 MTP	84.39%	0.829
Qwen3.5-397B-A17B-MXFP4 MTP	84.27%	0.828
MiniMax-M3-MXFP4 Eagle3	72.99%	0.715
MiMo-V2-Flash MTP	83.49%	0.82

*healthy pre-regression minimum. Run-to-run jitter is <1%, so 1.5pt headroom avoids false failures while catching real draft-head breaks.

Test

• python .github/scripts/validate_catalog.py → all catalogs OK.
• Workflow YAML parses; gate comparison logic verified (60.0% passes 0.6, 59.99% fails).

🤖 Generated with Claude Code

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 1557 --add-label <label>