[hsliuustc0106]

What

Third PR-analysis blog post: Understanding PR #6094 — Scheduler-managed diffusion KV cache control plane, following the format of #16 (DLO host-weight plans) and #20 (SVDQuant).

• blog/_posts/2026-08-16-pr-6094-diffusion-kv-control-plane.md
• blog/assets/figures/pr-6094-kv-control-plane/overview.png — the four-section design diagram (initialization chain · admission with atomic CFG allocation, backpressure and the deterministic never-fit gate · terminal frees + seq_len anatomy · coverage matrix), updated to the PR head c5f6159

Content notes

• Written from the review of the PR (both review rounds, including the author's follow-up commits at c5f6159 that addressed the open threads: deterministic never-fit block-sum, profile-envelope admission gate, ≤ 0 budget guard, rename).
• Numbers are author-reported from the PR body (dense A/B on 2×A100: identical 42.17 GiB, ~1% timing) and explicitly framed as no-regression evidence, not a perf claim; test counts (282 + 166) from the PR body.
• Links absolute; figure served from site.baseurl per the template rules.
• The same diagram is posted on the upstream PR (comment); the SVG source also lives in docs/visualizations/ on main (added earlier today — those raw URLs back the upstream comment, so they should stay).