Combine existing guarded common, Kimi and Muse source increments

[zozyo]

Unified source successor, 2026-09-22

This record does not update the frozen engine428 export, any Governor candidate,
image, CVM or production acceptance. No image build or registry write is needed.

The implementation is Phala-Network/sglang commit
6b3ca7eddd1f0a1eed2774bdfd461626c5e0e780, tree
b6730c9fe47e14319bb92dc61b8463ae7dcfedfd, on navigation branch
codex/model-union-v0520-20260922. It combines the three common increments,
six Kimi increments, and one Muse increment. No common patch bytes were copied.
Its complete tree equals the earlier combined candidate 83d8dc47ee; the new
linear history orders common first and shares the existing source commits.

unified-v0520-successor records the full source range, ten ordered patch
references, immutable commits and hashes. export_selectors.py independently
replays the 31-entry frozen series from official v0.5.20
94602c9c2b7cbdb8efd5c52802dac6a1c180089e to tree
83dcb00885129cc2afefdce2e169ebba697a9a67, then the successor to the tree above.
The source range is not a Governor v4 release input.

Current semantic scope

The original 110 per-source records, source hashes, ownership and old evidence
remain in COVERAGE.json. They describe frozen engine428 and
must not be silently relabeled as successor acceptance.

Model or family	Present in this successor	Still not established
Common	llguidance mask callbacks/pinning 87ace709c88e, typed usage 89f583819b2f, opt-in mode defaults 1db913dd7b38	Linux package/import and combined runtime tests
Kimi K3	Owner identity 1e62ad753c42, absolute checkpoint grid 9bd58a168358, prefill sequence cap 6329c4dddd08, file ownership/prefix cd0cca6903e8, chunk guards 5a1799c43d72, virtual page continuation 02558c2aaab1	TP8/DCP1/DCP8 runtime and new Governor combination; old CPU/CUDA evidence is scoped to its original source
Muse Glimmer	Required tool native channel delta 6b3ca7eddd1f; Gumbel/deferred-Mamba upstream coverage retains its original proof	Historical template/cardinality/grammar/cancellation deltas are not all proven equivalent by this one parser patch
Nemotron 3.5 Lightning	Frozen common behavior only	Literal/control-token source 4dbfdf98a4d5 does not clean-apply: serving-chat call wiring and reasoning-parser shared methods changed. The helper, request/token ID propagation, model/template guard and regressions must migrate together
DeepSeek V4.1 Flash	Frozen common behavior only	Native closure selector still fails at dsv41-0011-chat-encoding.patch, after ten successful applications. Complete native ABI/recipe and model protocol guards remain mandatory
Gemma 4 26B A4B	Common source; historical fork d0b3e70cbc5ed3cc757d22e79ffa1fff28f58571 remains identifiable	Shared old image does not prove a Gemma-specific increment or qualify Qwen GGUF paths
Qwen3.8 27B / Qwen3.5 GGUF	Current guarded Qwen/common implementation plus mode sampling defaults	Historical GGUF load, vision and Q8_0 source records still require semantic and hardware review
Qwen3.6 27B	Current common template, effort/history/schema/media paths; historical fork 711978779936d1918d038de8515e33f968cb5193 retained	Native XGrammar version/required/XML equivalence is not established by source export
Qwen2.5 7B	Common source only	Do not infer Qwen3.5/GGUF applicability from shared historical image
GLM 5.3	Existing frozen nine logical changes and formatting correction unchanged	No new GPU or Governor-on-GLM acceptance

Current source paths are the exact paths in each exported source delta.
Key boundaries are entrypoints/openai/serving_chat.py,
parser/reasoning_parser.py, function_call/muse_glimmer_detector.py,
managers/tokenizer_manager.py, managers/schedule_policy.py,
mem_cache/hicache_storage.py, and model_executor/runner/prefill_cuda_graph_runner.py
under python/sglang/srt.

Verification

• Four exact selector replays passed, including the combined successor.
• DeepSeek retained its expected failure; it did not become a pass.
• Reference-only selectors verify immutable identity but report passed: null.
• Seven real-Git fixture regressions passed: moved/deleted navigation refs,
  reference-only evidence, corrupt baseline bytes, wrong tree, unverified base,
  expected failure retention and invalid replay mode.
• No Linux runtime imports, GPU, final-image or production tests were run here.
  The local Python has no torch, tokenizers or transformers. Source/tree
  equivalence cannot replace those gates.

Run python -m unittest discover -s tests -v and
python scripts/export_selectors.py --source /path/to/sglang --check.
The main frozen export remains independently checked by scripts/export.py.
No additional per-model workflow or long-lived profile was added.