Three specs assert model_loader behaviour the loader no longer has, found while repairing their line citations

[localai-bot]

Found while converting model_loader.cpp line citations to symbol anchors for #1143. Three specs do not merely point at the wrong LINE — they assert loader behaviour the loader no longer has, so "repairing the anchor" would have parked a fresh-looking citation next to a false claim. They are recorded here instead of repaired, because each one belongs to its owning row's judgement, not to a citation sweep.

Verified against src/vllm/entrypoints/model_loader.cpp at fd64c76ee.

Spec	The claim	What the loader does now
.agents/specs/safetensors-windowed-load.md:63,108	"shards.clear() releases them after the whole load"; quotes the line shards.clear(); // the mmap'd shards may be released after the load.	shards.clear() does not exist anywhere in the file. The shard vector is a std::make_shared<const std::vector<SafetensorsFile>> whose last reference is held by the deferred-expert closure, released when that closure is (model_loader.cpp:1636-1643). The spec's "Referenced-live-after-load ranges: NONE" conclusion rests on the munmap-immediately behaviour it describes.
.agents/specs/gguf-dflash-draft.md:17-18	the three loader seams are "still typed on std::vector<SafetensorsFile>"	LoadDflashDraft takes a SharedHeadSource (model_loader.cpp::SharedHeadSource), which is exactly the B1 abstraction the same spec proposes further down as unlanded work.
.agents/specs/gguf-dflash-draft.md:373,380 (A5, B2)	a GGUF rejection to leave alone / to drop, at model_loader.cpp:717-723	717-723 is ResolveEnablePrefixCaching. The loader now #includes vllm/model_executor/models/qwen3_dflash_gguf.h and states at model_loader.cpp:906-910 that "the refusal is GONE rather than widened".
.agents/specs/model-factory-registry.md:91	model_loader.cpp:232-238 is IsDenseArch (num_experts==0)	232 is DirectDeviceLoadRequested. IsDenseArch survives only in a comment at model_loader.cpp:1515 saying the inline split "is superseded by" the registry.

Why this is one issue and not four fixes. Each row is a claim its owning row made and can re-derive; a sweep session can see that the claim is false but not what the true statement should be. #1143 is about a citation surface that cannot survive an edit; this is the second-order damage that surface hides — a reader who follows a stale line number lands somewhere plausible and never learns the sentence above it stopped being true.

Owner: none yet. Listed under ## Owed in citation-anchor-freshness.md until a row claims it.

[localai-bot]

Repair path and evidence: PR #1203 records this under ## Owed in .agents/specs/citation-anchor-freshness.md, which is where the owner comes from until a row claims it.

[localai-bot]

A fourth spec carries the same falsified anchor, found in the fresh review of #1203.

Spec	The claim	What the loader does now
.agents/specs/gguf-mtp-spec-decode.md:11	"Today LoadedEngine::FromModelDir refuses the combination outright (src/vllm/entrypoints/model_loader.cpp:717-723)", quoting the refusal message verbatim	The same removed GGUF refusal already recorded here for gguf-dflash-draft.md:373,380, and the same stale range: 717-723 is ResolveEnablePrefixCaching. The spec states it in the PRESENT tense, so a reader who follows the anchor lands on prefix caching and never learns the sentence above it stopped being true.

It is the same defect and the same owner, so it belongs on this issue rather than a new one. That makes it FOUR specs, not three; the title is left as written because an edited issue title loses its own history.

Recorded under ## Owed in citation-anchor-freshness.md.