Block PrefillDecode LiteRT spikes before dead-end loads

[Nostromo-618]

Summary

• Root cause: Ministral/Qwen3 .litertlm spikes are PrefillDecode models. Default LiteRT Backend.GPU_ARTISAN always uses streaming ModelAssets → Streaming kTfLitePrefillDecode models is not supported yet. Blob buffering does not help (runtime still streams). Probed Backend.GPU non-streaming createEngine path — still fails with null function on current @litert-lm/core web wasm. Chat sendMessage (non-stream) is irrelevant because load never succeeds.
• Product fix: Mark spikes litertRuntime: 'prefilldecode-unsupported', block AiChat.load() before download, disable Load with “Runtime unsupported”, show a clear pre-load reason (not just Experimental), keep spikes in the catalog for honesty.
• UI: Stop duplicating the same error in muted progress text + red banner.
• Docs/tests: Update spike notes; add unit coverage for PrefillDecode detection, error rewrite, and load block.

Test plan

• pnpm test -- tests/unit/guardrails.spec.ts (includes PrefillDecode block test)
• Manual: open AI Chat, select Ministral 3 3B LiteRT — Load button shows Runtime unsupported; no download; single clear message (no red+black duplicate)
• Manual: select Qwen3 0.6B LiteRT spike — same blocked behavior
• Manual: select Gemma 4 E2B — Load still works
• Manual: select WebLLM Tiny (Qwen3 0.6B MLC) — Load still works

Made with Cursor

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Advisory notes (legacy AiChatUI, not in diff — no inline comments):

1. _renderCapacityNote (ai-chat.js:~2439) does not suppress capacity warnings for loadBlocked models, unlike the Vue component's refreshCapacityNote. A blocked spike could show a misleading "this device may struggle" note alongside the Runtime unsupported banner.
2. _renderCacheHint (ai-chat.js:~2204) called for blocked models without clearing the hint text — shows "First load downloads model files from the web" even though load is impossible. The Vue component clears cacheHint for blocked models.
3. _handleSwitchModel (ai-chat.js:~2551) lacks the loadBlocked early-return that _updateSwitchButtonState already enforces at the button level — a defense-in-depth gap. The Vue component's switchModel has this guard.

The legacy AiChatUI is labeled compat/test-only, so these are low-risk, but worth addressing for consistency.

Files Reviewed (5 files)

• README.md
• ai-chat.js - new functions, guards, and legacy UI error dedup
• doc/vdl-ai-chat.md - spike notes updated
• src/components/VdlAiChatUI.vue - loadBlocked state, error dedup, UI changes
• tests/unit/guardrails.spec.ts - PrefillDecode block test

---

_{Reviewed by deepseek-v4-pro · Input: 82.6K · Output: 18K · Cached: 803.2K}