docs(sglang): record the MTP penalty-accounting defect we filed but do not patch

[dorado269]

What

One row added to deploy/docker/patch.upstream.status.md for an sglang defect I filed upstream as sglang#33643 (ours, dorado269, DRAFT) and deliberately did not patch locally.

The defect: a decode step commits accept_len tokens under MTP, but cumulate_penalty_output_tokens feeds only output_ids[-1] to the sampling penalizers. Every other accepted token never enters penalty state and is never backfilled, so frequency_penalty ends up effectively divided by the accepted length, and min_new_tokens counts decode steps instead of tokens — masking EOS to -inf for roughly accept_len× longer than the request asked.

Why it needs a new section

The page opens with "Every patch under deploy/docker/patches/", and every section is a patch directory. This row has no patch file, so there is no patch header to hold the argument the way ## Maintenance assumes. Rather than bend an existing table, this adds ## Filed upstream, not patched here before ## Not patches, and says in the intro and in ## Maintenance that the section runs the other way.

The point of the row is that "we know, and here is why we are not carrying a diff" should be answerable from this page — otherwise the next person who finds #33643 re-argues it.

Why no local patch

The path is gated. eagle_prepare_for_decode calls cumulate_penalty_output_tokens() only when penalizer_orchestrator.is_required, which needs some in-flight request to carry a non-default frequency_penalty, presence_penalty, repetition_penalty or min_new_tokens. Nothing in this repo sets one — grepped 2026-08-05, the only penalty hits are unrelated prose about load and seek penalties.

The non-obvious way it could turn on is the model rather than our configs: --sampling-defaults defaults to model, so sglang copies repetition_penalty out of the served weights' generation_config.json into every chat request's defaults. Upstream zai-org/GLM-5.2 carries only temperature and top_p there, read from the raw file rather than inferred from a search, so the gate stays shut. We serve the MXFP4 derivative and it was not on the host to check — that caveat is stated in the row rather than glossed.

Verification

claim	how
nothing in this repo sets a penalty	grep across the tree, 2026-08-05
zai-org/GLM-5.2 has no repetition_penalty default	raw generation_config.json fetched 2026-08-05
--sampling-defaults defaults to model	read server_args.py in sglang main
#28180 and #29512 still OPEN, titles as quoted	gh issue view, 2026-08-05
MXFP4 weights' generation_config.json	not checked — weights not on the host, and the row says so

Docs only, no code or Dockerfile touched. Draft because #33643 is itself a draft; the PR state column should be refreshed if that changes before this merges.