[qwzx-qwas]

Description

Follow up on the nonblocking review leftovers from merged #862. The model-loader wrapper has no remaining caller after the scheduler tests moved into the crate, and the HF gate constructs four invalid targets that the production selector cannot emit. Remove that wrapper and redundant loop, narrow recipe storage/preparation visibility, and retain the actual-model selection assertions, production rejection checks and numerical thresholds.

Remove retired identity-interface explanations and the private Stage 20 message. Add a short Quick Start covering the separate pinned generation environment, independent FlashInfer checkout, bundle generation, release build, explicit candidate selection and HTTP verification. Runtime algorithms, defaults, AOT contracts and shared tuning remain unchanged.

Related to #691; this cleanup does not complete that migration or close the issue. No private notes, delivery archives, generated objects, caches or raw results are included.

Type of Change

• Refactoring / review cleanup
• Documentation update
• New feature

Validation

In response to the reviewer's request to run Clippy while hosted CI is having problems, the retained CUDA environment completed:

cargo build --release --locked --offline \
  -p pegainfer-server -p pegainfer-qwen35 -p pegainfer-kernels \
  --all-targets --no-default-features --features qwen35

cargo clippy --release --locked --offline \
  -p pegainfer-server -p pegainfer-qwen35 -p pegainfer-kernels \
  --all-targets --no-default-features --features qwen35 -- -D warnings

Build passed in 1m20s; Clippy passed in 39.89s, using nightly-2026-07-10 and CUDA 13.1.115 with the existing candidate linked. The returned evidence archive and internal hashes were verified, and its source manifest matches every current source/test file; only this PR's README prose was extended afterwards. The dependency future-compatibility notice for proc-macro-error2 remains in the log, with no compiler or Clippy error.

Rustfmt, Bash syntax, local ShellCheck and git diff --check passed. The build includes test targets but does not execute them. HF/HTTP/ABBA were not repeated for these removals, and the earlier #862 GPU results are not presented as a new run. Hosted CI on this follow-up remains to be checked.

Checklist

• Changes reviewed against the remaining feat(qwen35): add opt-in SM120 FlashInfer GDN prefill #862 comments.
• Relevant compilation, strict Clippy, format and script checks passed.
• Commitizen commit with matching Gmail author/committer and DCO sign-off.
• Only the four reviewed files are included; no private/generated artifacts.
• Hosted CI on this PR head has completed successfully.

[chatgpt-codex-connector]

Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-08T11:19:28.812254Z	b0391b7	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.