fix(middleware): drain websocket session end streams

[pimlock]

Summary

Prevent WebSocket middleware session-end notifications from being canceled with their bidirectional gRPC stream. The supervisor now half-closes the request side and briefly drains the response side before releasing the transport.

Related Issue

No issue required: localized lifecycle fix found while investigating the Rust test hang in #3085.

Changes

• keep the tonic response stream alive while the terminal request leaves the local transport
• use the same bounded close path for preflight exits, explicit session ends, disabled stages, and dropped sessions
• bound the skip regression assertion so delivery failures report instead of hanging the test process
• document the stream shutdown contract for middleware implementations

Before and after

sequenceDiagram
    participant S as Supervisor
    participant H as tonic / h2
    participant M as Middleware

    Note over S,M: Previous shutdown
    S->>H: Queue session_end
    S-xH: Drop request and response handles
    H--xM: RST_STREAM(CANCEL) may discard session_end

    Note over S,M: Graceful shutdown
    S->>H: Queue session_end
    S->>H: Drop request sender to half-close
    H->>M: Deliver session_end, then request EOF
    M-->>H: Finish response stream
    H-->>S: Drain completes or reaches the 10 ms bound

Loading

Testing

• mise run pre-commit passes
• cargo test --workspace --exclude openshell-server passes
• cargo test -p openshell-server --features test-support passes
• all 85 openshell-supervisor-middleware tests pass
• all 66 supervisor-network WebSocket tests pass against h2 0.4.19
• E2E tests added or updated, not applicable for this localized transport lifecycle fix

Checklist

• Follows Conventional Commits
• Commits are signed off for DCO

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-3143.docs.buildwithfern.com/openshell