Re-evaluate DFlash 2 speculative decoding when Ollama or llama.cpp ships it

[atomantic]

Context

DFlash 2 speculative decoding was evaluated on 2026-08-19 and rejected for now — full write-up with the verified evidence in
docs/research/2026-08-19-dflash2-speculative-decoding.md.

Short version: DFlash 2 is engine-side decoding machinery plus a drafter checkpoint. Ollama — the backend that serves the Ollama
provider and both Ollama-backed TUI providers (Claude, OpenCode) — ships DFlash v1 (x/models/dflash), has no DFlash 2
candidate-selector / depthwise-conv modules, and publishes no -dflash tag for the Qwen3.8 targets. Every engine that can run it
(llama.cpp, vLLM, SGLang) is reachable only from an unmerged PR, and adopting one would mean PortOS taking on a third local
inference backend built from a patch.

Trigger to re-open

Re-evaluate when either holds — either one collapses the work to a catalog entry rather than a backend:

1. Ollama publishes a packaged DFlash 2 tag for a Qwen target — a qwen3.8:27b-dflash2-shaped tag resolving to one
   self-contained model — and it measurably decodes faster than qwen3.8:27b-mlx on the same machine.
2. llama.cpp #27342 merges and Ollama's vendored engine picks it up.

Note that Ollama's existing DFlash v1 tag on another model was reported to decode at non-DFlash speed
(ollama#17683, closed not planned) — so "a tag exists" is not sufficient. Benchmark
before cataloging.

Decision already made (do not re-litigate)

If the trigger fires, the integration is a curated catalog entry for the packaged tag, not a new backend and not a drafter
sidecar install. This follows the precedent in docs/research/2026-08-16-qwen38-mlx-macos.md: PortOS recommends the complete,
engine-native model and does not own a speculative target/drafter pairing runtime.

Work when unblocked

• Benchmark the packaged tag against qwen3.8:27b-mlx on Apple Silicon (tokens/sec, decode only, same prompt and context).
• If it wins: add the tag to the curated local-LLM catalog (server/lib/localLlmCatalog.js), gated to the platforms where it is
  actually faster, mirroring how qwen3.8:27b-mlx is gated to Apple Silicon.
• Re-check the drafter exclusion in server/services/huggingFaceCatalog.js (hasDrafterTag / isMlxDrafter) — a packaged tag
  must not be filtered out by it, while the bare drafter sidecars must stay filtered.
• Add a short "superseded by" note at the top of docs/research/2026-08-19-dflash2-speculative-decoding.md.

Acceptance criteria

• Either the packaged tag is cataloged with a benchmark number recorded in a dated research note, or this issue is closed with the
  benchmark showing no win.
• No drafter sidecar becomes installable as a standalone chat model as a side effect.

Blocked on

Upstream: llama.cpp #27342 and/or an Ollama release shipping a DFlash 2 Qwen tag. Labeled blocked so autonomous claim runs skip it
until then.

[atomantic]

Closed as resolved by PR #4570 (OpenCode llama TUI provider and DFlash 2 speculative decoding integration shipped in v2.46.0).