PLAN: Vulkan through the apps on three GPUs, and the 16k MediaTek chunk rule

[alpharomercoma]

Documentation only: no pipeline code changes.

• Out-of-date statements fixed: docs/PLAN.md said the apps can't load Vulkan files and pinned ExecuTorch 1.4.0. Both changed on 2026-10-04; the Vulkan row in the phase table now reads done.
• Finding 38 (Vulkan on the phone): Qwen3-0.6B 2k, run through openweights and ExecuServe on a Mali-G925, an Adreno 750 and the SM8850's Adreno.
  • Correct on every GPU.
  • On Mali the CPU file decodes about 3× faster.
  • On the SM8850 the GPU reads long prompts 1.4–1.9× as fast, but decodes short replies at 0.7× the CPU's speed.
  • The study and its raw logs are in openweights docs/research/vulkan-on-device.md and tools/eval/results/vulkan-2026-10-04/.
• Finding 39 (MediaTek 16k): LFM2.5 at 16k built only once each chunk held at most one attention layer: 8 chunks for the 1.2B, 10 for the 2.6B. The 32k runs were cancelled for budget, so it's not known whether they build.
  • The max_chunks input's description in export-mtk.yml now says this.
• CLAUDE.md: the local test pins follow the 1.5.1 bump.

ruff check and pytest pass (205 passed, 4 skipped).

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is enabled.}