dots3-note: the row umbrella #699 is a deleted issue, and 36 records point at it

Row: MODEL-MM-dots3-note-dots3-note-for-causal-lm

#699 — the issue every ## Owed entry on this row names as its tracking
target — does not exist. It is a 404 to an authenticated caller and to an
anonymous one alike, while #697 and #2881 answer 200 to both, so this is a
real deletion and not the read-side failure an account restriction produces.
#698, #700 and #701 are 404 the same way: a contiguous deleted block.
I could not determine whether they were deleted, transferred, or never filed.

This issue replaces it as the row's umbrella, so the owed work has a live
target to point at.

What points at the dead number

36 citations in the issues/699) URL form across nine files:

Plus two bare #699 references in docs/FEATURES.md (:192, :257) and one
in a scripts/check-agent-record.py:121 comment.

Not all of them should move. No checker's logic reads the number — the
check-agent-record.py mention is a comment explaining how the model ratchet
reached 377 — so this is a broken-pointer defect, not a gate failure. A dead
link inside a historical justification is evidence of what was true when it was
written and stays put per AGENTS.md §Records. What must move is every LIVE
pointer: the Owner: this row. Issue #699 field on each ## Owed entry, the
row's matrix cells, and the two docs/FEATURES.md cells that offer it as the
place to follow the gap. The citations in glm-dsa-latest-deepseek.md,
kv-dsv4-multicache.md and 2026-09-03-portq4.md belong to other rows and are
theirs to repoint, not this row's.

Two record claims the tree has already falsified

Found while scoping W10, both verified against the tree at d023e3357:

1. The ## Owed nextn entry names a refusal that no longer exists. It says
   "W4a refuses a config with num_nextn_predict_layers > 0", and W5c
   (#2176) deleted that refusal. Dots3NoteDeviceRefusal now returns "" for
   the released config (src/vllm/model_executor/models/dots3_note_device.cpp:872),
   its nextn branch survives only as a comment block at :851-868 headed
   "LIFTED at W5c", and tests/vllm/models/test_dots3_note_attn.cpp:3500-3504
   asserts the emptiness with num_nextn_predict_layers = 1. The same spec says
   so at :8494-8498 and contradicts itself at :8996-9001. The live predicate
   is now the deferral Dots3NoteIsNextnTensor
   (src/vllm/model_executor/models/dots3_note.cpp:558-572), which is a
   different shape with a different scope boundary.

2. The layer_types reconciliation W1/W2 recorded as open is answered
   upstream. The spec says config.layer_types has no entry at the nextn
   index and that upstream cannot settle the sliding half. It can:
   SpeculativeConfig.hf_config_override extends layer_types with
   ["sliding_attention"] * n_predict before the decoder layer indexes it
   (vllm/config/speculative.py:648-666@3b45d053b4), and there is an override
   key mtp_layer_types the released config does not carry. W1/W2 read only
   model.py and so missed it. The fixture agrees: layer_types holds exactly
   46 entries against num_hidden_layers = 46, so without that override index
   46 is an IndexError rather than an ambiguity.

Both are BEYOND-PIN reads (3b45d053b4, against pin 5559679229), consistent
with this row's §6.1.

Why this is not folded into a brick

AGENTS.md §"Work happens in a worktree" makes a record-only change a legitimate
unit of work when the record IS the work — "a stale row, a corrected pin, a
newly filed gap" — and all three of these are exactly that. None of them
restates what a landing change already knew, which is the case that rule
excludes.

Owed to: this row.

Start with AGENTS.md §Records, then inspect the listed .agents files, docs/FEATURES.md, and the Dots3Note spec references around the cited source and test locations. Repoint only live pointers to issue #2924, preserve historical citations, and reconcile the two falsified record claims while keeping the documented scope boundaries intact.