Nothing to Revoke: Canvas Allocation, Not Commitment Revision, Governs Code Generation in a Production Diffusion Language Model

Decoding research for diffusion language models increasingly targets the reverse process by revoking committed tokens, gating commitments, and grouping acceptance by syntax. We measure what these levers are worth on DiffusionGemma-26B, whose transient acceptance makes every token revocable at every step for free, using exact replay, common random numbers, pre-registered decision rules, and the official BigCodeBench evaluator. The reverse process has nothing to act on. Commitment regret is 0.12\% over 2.6 million events, a tenfold sweep of the entropy bound changes nothing, 87\% of final tokens are fixed by step 5 of 16, and six reverse-sampling policies at matched compute are indistinguishable from stock, the two largest-scale within a point on 707 tasks. The leverage is elsewhere. Under a fixed single-canvas budget, block diffusion spends 55\% of its canvas on prose and truncates 96\% of its programs; reallocating the canvas to code with one instruction raises Pass@1 by 33.6 points (95\% CI [30.2, 37.1]) on 707 tasks at the same 16 forwards; a second canvas recovers most of this gain at 17.9 forwards, so allocation is the cheaper fix. The self-conditioning channel, the model's only inter-step state, governs wording rather than content, so errors are decided as beliefs before commitment. Where compute pays, it pays through allocation. Roughly 20 to 27\% of hard failures are recoverable by resampling, majority voting matches an oracle when the correct answer is modal, and recoverability is predictable from one trajectory.