refactor(connector): restore pipeline behavior and boundaries

[GentleCold]

Summary

• Keep SchedulerConnectorMixin and WorkerConnectorMixin as separate vLLM role adapters.
• Organize scheduler implementation under daser/connector/scheduler/ and worker implementation under daser/connector/worker/.
• Restore bounded concurrent store execution, streaming buffer reuse for large requests, independent load failure handling, and shutdown cleanup.
• Restore request-level load performance diagnostics and publish L1 hit, L1 miss, and L2 read metrics.
• Derive independent fixed load/store pools from one explicit 6 GiB CUDA staging ceiling per worker.
• Make DaseR drain failures abort benchmarks and require both L1 and L2 activity in evict-mode warm validation.

Architecture

DaserConnector remains the vLLM composition root. Scheduler and worker implementation packages do not import each other.

• RequestLifecycle owns scheduler request state and synchronous control-plane IPC.
• WorkerRuntime owns finalized KV geometry, step metadata, pipeline configuration, and coordinated shutdown.
• LoadPipeline owns load queueing, fixed staging leases, IPC, restore, diagnostics, failure invalidation, and completion.
• StorePipeline owns deferred saves, producer ordering, fixed staging leases, transfer, commit, and completion.
• Shared metadata and IPC contracts remain at the connector root.

No pipeline ABC, generic dispatcher, compatibility re-export layer, or variable-size CUDA allocator was added.

Behavior restored

• Independent store requests execute concurrently up to store pool depth.
• Requests larger than the pool stream batches sequentially through one reusable buffer.
• Load requests complete and fail independently; failures invalidate only their affected block IDs.
• Restored KV becomes visible after the dedicated load stream synchronizes, without device-wide synchronization.
• Existing transfer responses now provide request-level bytes, copies, copy runs, IPC time, dispatcher wait, KV copy, worker sync, transfer open/load/sync, L1 hit/miss, and L2 read diagnostics.

Staging memory

Load and store retain independent fixed pools so store traffic cannot consume load buffers. One startup calculation enforces a combined ceiling of 6144 MiB per worker and reserves at least one buffer for each direction.

The TP2 validation run selected load depth 3 and store depth 2:

Worker	Combined allocation
TP0	5,921,570,815 bytes
TP1	6,010,699,775 bytes

Both are below the 6 GiB ceiling (6,442,450,944 bytes).

Validation

• pre-commit run --all-files
• pytest -q tests/connector (119 passed)
• pytest -q tests/server tests/transfer (154 passed)
• benchmark utility tests (106 passed)
• pytest -q -m "not integration and not slow" (353 passed, 100 deselected)
• Complete no-evict master and PR benchmarks
• Complete PR evict benchmark with successful drain, no failed requests, and both cache tiers exercised
• Benchmark scratch data and service processes cleaned up

Performance parity

Metric	Master	PR	Difference
Cold mean TTFT	8,630.27 ms	8,571.86 ms	-0.68%
Cold prompt throughput	17,893.93 tok/s	18,015.26 tok/s	+0.68%
Warm mean TTFT	441.46 ms	439.92 ms	-0.35%
Warm prompt throughput	349,735.13 tok/s	352,416.58 tok/s	+0.77%
Warm backend hit	100%	100%	unchanged
Failed requests	0	0	unchanged

The PR evict run completed with l1_hits=176, l1_misses=880, and l2_reads=880. Its complete warm phase mean TTFT was 553.52 ms and prompt throughput was 278,552.04 tok/s.

Cold/warm exact match was 98% on both master and the PR, so the remaining mismatch predates this refactor.

mypy . remains blocked by the repository's existing duplicate-module discovery (config and daser.config) and pre-existing annotation errors outside this change.