[Manual Test] [Core] prompt_preprocess_func in stage config (upstream #2796)

[lishunyang12]

Fresh test PR after tightening the claude-quick preview prompt.

Mirrors upstream vllm-project#2796 — [Core] Add prompt_preprocess_func to stage config (5 files / 208 lines). Uses the [Core] title prefix so the bot's Phase 1 will trigger the Core-specific checks (scheduler invariants, thread safety, high bar for tests).

Try the three tiers in order

1. claude-quick (Haiku) — should now give you a short signal-only preview: domain summary + risks + depth rec. NO file-by-file lists, NO meta fluff.
2. claude-review (Sonnet) — full Phase 0–3 maintainer review. [Core] prefix should trigger scheduler/thread-safety checks in Phase 1.
3. claude-deep (Opus) — same scope as Add claude review #2, stronger reasoning for subtle concerns.

Compare the three outputs. The Haiku preview should be short and straight — if it still pads with file lists or meta reciting, the prompt isn't tight enough. Do not merge.

[claude]

Preview (Haiku)

What it does: Adds server-side prompt preprocessing hook that transforms raw user prompts before Stage 0 tokenization, enabling chat template or system prompt formatting without client-side changes.

Risk flags: Modifies critical request initialization path. Test coverage gaps: no integration test of the actual preprocessor execution during request creation, only metadata loading/collection. [Core] prefix implies scheduler/thread-safety checks — concurrent requests must safely invoke the preprocessor function.

Suggested depth: claude-review — [Core] PR touching engine request path needs Phase 1 thread-safety verification and test audit for execution path coverage.

[lishunyang12]

Closing — tone + link updates applied. New test PR follows.