refactor(adapters): simplify vLLM and SGLang ownership

[feichai0017]

The adapters loaded native/GPU modules during plugin discovery and retained an unused vLLM availability owner and health thread. This cleanup loads only the selected vLLM role, closes native connections after initialization failure, uses the pinned engine's unchanged optional callback defaults, and reports registration conflicts explicitly. SGLang graph-event ownership moves into events.py, with unselected backend hooks remaining lightweight. Existing page-retention, recovery and P/D lifecycle owners remain in place.

This PR targets the accepted feat/s2-metadata-qualification baseline so its diff contains only the two adapter-cleanup commits (61b2ac0e, 13bd405a). The adapter guide records the pinned LMCache/FlexKV references and ownership boundaries. No native ABI or Rust implementation changes are included.

Validation:

• Source-only Python: 414 passed, 1 skipped; pinned native contract/admission gates: vLLM 7 passed, SGLang 22 passed.
• A100, Qwen3-8B, TP=1/PP=1: vLLM 0.29.0 serving gate 6 passed, 1 hybrid-only assertion skipped; SGLang 0.5.20 DRAM and io_uring SSD recovery passed, including restart and cold controls.
• SGLang event/linker integration: 19 passed, covering real GPU page overwrite and hybrid-state recovery plus controlled failure/drain cases.
• Scoped prek checks pass; Astro check has no errors/warnings; final website build produces 44 pages and both publication tests pass.
• 130 Python files verified, 7 reused frozen native artifacts unchanged, and recorded Manager sockets/processes cleaned up.

Two initial SSD starts exposed a test-harness port bug: loopback availability did not imply wildcard availability for TCPStore. A standalone reproduction confirmed it; the helper now checks wildcard listening and has a real-socket regression test. The fixture now preserves Manager logs. Failed-run evidence is retained; the original fixture had already deleted its first-run temporary Manager logs.

External reviewer handoff: /root/orbitkv-artifacts/adapter-cleanup-20260930/HANDOFF.txt. Downloaded logs and hash manifest: remote-evidence/ in that directory. Full A100 runs: /workspace/orbitkv-adapter-cleanup-20260930/. Qualification and remaining work are recorded in completion-plan S5.2; independent review, engine upgrades and broader multi-GPU/P/D qualification remain open.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Advanced

Run ID: 2cb4c3d2-075c-4aa6-8e04-bca50243160d

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}