162 of 239 issue citations in shipped docs/ are 404, including both on the docs/FEATURES.md streaming row

[localai-org-maint-bot]

Row: ENG-RECORD-ANCHOR-RATCHET

Named on that row because it owns the class "the record's citations were range-checked
and never reported". This is the same class one surface over: citations that no longer
resolve, URL-shaped rather than path:line-shaped, in the SHIPPED docs rather than in
.agents/. Re-home it if the row owner disagrees; it needs an owner more than it needs
that particular one.

What

162 of the 239 distinct issue and pull-request numbers cited in docs/ and README.md
return 404. That is 68% of the citations on the surface users read.

Measured at origin/main 23ac6f1a7, by extracting every
github.com/mudler/vllm.cpp/(issues|pull)/N from docs/ and README.md, deduplicating,
and asking gh api repos/mudler/vllm.cpp/issues/N for each. A dead number returns
Could not resolve to an issue or pull request — not a closed state. These are gone,
not resolved.

Where, worst first

 42/ 79 dead  docs/FEATURES.md
 37/ 44 dead  docs/ENVIRONMENT.md
 27/ 50 dead  docs/USAGE.md
 19/ 27 dead  docs/benchmarks/open-gaps.md
 10/ 11 dead  docs/benchmarks/vllm-online-serving.md
  9/ 10 dead  docs/benchmarks/at-a-glance.md
  8/  9 dead  docs/guides/multimodal-input.md
  8/ 11 dead  docs/SPECULATIVE-DECODING.md
  7/  8 dead  docs/benchmarks/memory.md
  6/  7 dead  docs/models/qwen3-8-2-4t.md
  6/  7 dead  docs/guides/expert-streaming.md
  5/  6 dead  docs/models/qwen3-8-27b.md
  1/  4 dead  README.md

(38 further files carry between 0 and 4 dead references each.)

The cutoff is clean, and it says what happened

Every docs/bench-evidence/*-2026090[23]*.md file scores 0 dead. Everything written
earlier is mostly dead. The surviving population starts around #1870. This is not gradual
link rot; it is a single loss event below a number.

Why it is worth a row rather than a sweep

Two of these are load-bearing on the front page of the feature table.
docs/FEATURES.md:64 is one row of the KV-cache table and carries both:

| Routed-expert weight streaming from disk | ◐ default OFF (VT_MOE_EXPERT_STREAM=1), keep-quant/keep-f16 towers (#1378); bounded slot cache; refuses unfittable slices by name. c1-c4 capacity, not throughput. CPU; staging device DECODES, token gate FAILS (#1299) |

Both citations are doing real work and both are dead:

• The expert-stream lane fires without checking the residency ROUTE, deleting the #1123 refusal on the non-keep-quant arms #1378 supports the scope restriction "keep-quant/keep-f16 towers". It is still
  TRUE of the code — GgufExpertTowersReachSlotLane
  (src/vllm/model_executor/model_loader/gguf_device_fit.cpp) accepts kKeepQuant and
  kKeepF16 and returns false otherwise — so a user who follows the link to understand
  why their NVFP4 GGUF gets no streaming lands on a 404.
• Qwen3.8-2.4T UD-Q1_0 LOADS on --device cuda and then exhausts a GB10 in its first forward: the non-expert weights are resident twice #1299 supports "staging device DECODES, token gate FAILS". Also still true: the
  row spec records G0-CORRECT as FAIL and G0-SPEED as VOID. A user reading a failing
  correctness gate on a shipped feature cannot reach the evidence for it.

So the docs are accurate and unverifiable at the same time, which is the worst of the two.

Not a decision this issue makes

Three repair shapes, and picking one is the work:

1. Re-file replacements for the load-bearing numbers only (as Device expert slot store: ExpertSlotStore still has no virtual read, so the seam holds the concrete host type (replaces dead #1124) #2824 does for Qwen3.8-2.4T on --device cuda needs a DEVICE expert slot store: the larger-than-memory lane is host-only by construction #1124) and
   re-point those citations. Highest fidelity, most effort, and it renumbers history.
2. Strip the hyperlink and keep the bare number as a historical marker, with one note per
   document saying the tracker lost everything below VT_GGUF_KEEP_QUANT=0 is unreachable on a 16 GiB ROCm card: the bf16 fallback OOMs with an allocator throw, not a refusal that names the cause #1870. Cheap, honest, loses the
   evidence trail.
3. Point the dead ones at the recovered text in .agents/completed/issue-index.md, which
   still carries a detailed row for most of them. Keeps the evidence, adds an indirection.

The recovery source is real and rich: .agents/completed/issue-index.md is 913 rows, each
carrying the full description with file:line anchors and measured numbers, and
git log -p --follow -- .agents/issue-index.md covers 1659 commits for anything the final
version dropped.

Caution for whoever takes it

The recovered text is weeks stale and must be re-grounded before it is re-published.
Grounding #1124 for #2824 found two of its four claims already fixed, and a separate pass
found #1126's central claim ("CudaBackend never overrides Backend::DeviceMemoryInfo")
flatly falsified by src/vt/cuda/cuda_backend.cu:93. Copying a recovered row into a live
document without checking it would republish finished work as owed.