Oracle policy has no fallback and no pin concept: when vLLM has no implementation, secondary oracles (transformers, diffusers, SGLang/SGLang-Omni, llama.cpp) are used unpinned and unenforced

[localai-bot]

The gap

AGENTS.md §"vLLM is the reference" says comparisons run against the pinned
oracle in .agents/upstream-sync.md. That surface pins exactly one thing: vLLM.

But work in this repo already gates against four other upstreams, and every one
of them is pinned in a different spec file, or not at all:

Upstream	Where its pin lives today	Registered?
vLLM	.agents/upstream-sync.md parity-pin block	yes
vLLM-Omni	nowhere — #633	no
SGLang	.agents/specs/cuda-sglang-low-concurrency.md (v0.5.13 / 28b095c0)	no
llama.cpp	~6 CPU specs (237ad9b96 / b9892)	no
HF transformers	only as a component of the vLLM env line	no
diffusers	nowhere	no
SGLang-Omni	nowhere	no

Consequences already observed:

• vLLM-Omni has no parity pin: H3 W3+, LTX-2.5 and ~40 omni-only architectures (the whole TTS family included) cannot be gated against any oracle #633 — vLLM-Omni has no pin, so H3 W3+, LTX-2.5 and ~40 omni-only
  architectures cannot be gated against any oracle.
• A model whose ONLY upstream implementation is outside vLLM (e.g.
  MiniMaxAI/MiniMax-Music3, served by SGLang-Omni / diffusers, absent from
  vLLM) has no admissible oracle at all under the rule as written, so scoping it
  stalls on a policy question rather than an engineering one.
• A pin recorded inside one spec is invisible to every other spec, which is how
  the same upstream ends up compared at two revisions.

What this issue owes

1. AGENTS.md: state the fallback rule — vLLM (incl. vLLM-Omni) is the primary
   reference; where it has no implementation, a named secondary oracle from
   an admissible set is used, never an unnamed one, and never in preference to
   vLLM where vLLM implements the behavior.
2. A pin concept for secondary oracles, on a record surface that is one file
   per oracle (globbed for reading) rather than a shared file every PR must
   write.
3. Enforcement: a checker with a --self-test corpus and a mutation test,
   wired into scripts/agent-preflight.sh and CI — so the rule cannot rot into
   prose nobody runs.

[localai-bot]

Reopening: the policy, the registry and the checker landed in #650 (b81a5e41),
but this issue is also the owing issue that three registry records point at.
.agents/oracles/sglang.md, diffusers.md and tt-forge.md each carry
evidence = #647, and check-oracle-pins.py requires a gateable = no record
to name an open debt — so closing this would leave three records pointing at a
closed issue.

What remains owed here, one measurement each:

Oracle	What would make it gateable = yes
sglang	run SGLang (and SGLang-Omni) on this hardware and record a model that builds and runs. Source has been read at v0.5.15/f63458b5; nothing has been executed — its parity spec is explicitly "no measurement taken here"
diffusers	execute it as a scheduler/VAE oracle once, and record the resolved version. Currently UNPINNED because there is no measured revision to pin
tt-forge	needs Wormhole/Blackhole silicon. The BACKEND-TENSTORRENT row's real-hardware P150 is where this first becomes measurable; also has no release tags, so its pin will be a commit

vllm-omni is deliberately NOT listed: its pin debt is #633 and its record
points there.

Closing this issue is correct once all three carry measured evidence, or once
each remaining debt has been moved to its own issue and the records repointed.