The DFlash2 runner fixture drafts a CONSTANT, so every drafted-token comparison through it is a tautology against a numerics change

[localai-bot]

Found while landing SPEC-DFLASH2 W11 (#1890), by mutation rather than by reading.

tests/vllm/v1/spec_decode/dflash2_runner_fixture.h drives the production DFlash2 runner and RunAndCollectDrafts returns the drafted blocks off the VT_SPEC_TRACE line. Every one of the eight steps drafts the same three tokens:

threw=[] n=8
  [0] 19 19 19
  [1] 19 19 19
  [2] 19 19 19
  [3] 19 19 19
  [4] 19 19 19
  [5] 19 19 19
  [6] 19 19 19
  [7] 19 19 19

Measured on main @ 1724be38e + the W11 branch, Release, CPU, with a one-off probe that printed what DraftedBlocks returns.

Consequence. Any case that compares two arms by these strings compares one constant against another, so it cannot see a numerics change in the draft forward. Three mutations of the W11 draft-block attention, each a genuinely WRONG attention, left the comparison green:

mutation	what it breaks	drafted-token comparison
force the layer mask causal when the layer is non-causal	the whole block mask	green
neutralise the paged K/V write before the read	the draft attends over unwritten pool slots	green
hand the read the store's seq_lens instead of the extended bound	every block row silently dropped	green

The byte-for-byte op-level gate W11 added (test_qwen3_dflash_block_route) reds on all of them, which is how the degeneracy was located.

This is not only a W11 concern. dflash2 runner (W8): the paged lane and the materialized lane draft identically compares the same constants, so its stated guarantee — that two attention lanes agree — is not measured by it either. The W9/W10 cases assert counters and trace presence rather than token values and are unaffected.

Not the D9-scalars case. dflash2 runner: D9's SCALARS move the drafted tokens, in production DOES observe a difference, so the fixture is not degenerate under every perturbation — only under changes that move the block logits by less than whatever separates token 19 from its runner-up.

What a fix looks like: make the synthetic draft weights produce a non-constant argmax across the block (a wider draft vocab, or weights whose per-position logits actually separate), and then re-state the W8 and W11 comparisons as the numerics gates they are meant to be. It has to be done without disturbing the W3/W4/W9/W10 cases that read the same fixture.

NOT FIXED IN FLOW: changing the fixture's weights moves five landed cases at once and needs its own red-before evidence, which is a different unit of work from #1890. Owned by row SPEC-DFLASH2 and listed under ## Owed in dflash2-draft-block-fa2.md.

[localai-bot]

Independently reproduced from the other end while landing #2010 (row/FIX-DFLASH2-CONCURRENCY).

Mutation B' on that branch — delete the propose_drafts_block position invariant at src/vllm/v1/worker/gpu/runner.cpp:2939-2945 — was NOT caught by the focused gate, because the block-comparison leg that would have caught it is a tautology on this fixture for exactly the reason this issue records. The author wrote that leg, ran it, saw nine passing assertions measuring nothing, and removed it; that removal is what surfaced the ungated invariant.

So this fixture's constant drafts are not only blind to numerics changes (as filed) but also blind to deletion of a runner-side correctness guard. Filed separately as #2009 before the duplication was noticed; #2009 is now closed in favour of this issue.