test(hicache): cover buffer pipeline idle flush gate

[vuuihc]

Motivation

Buffer-only HiCache must not report the scheduler as fully idle while D2H or H2D pipeline work is still in flight. Otherwise, flush_cache() can reset cache state before transfer acknowledgements and staging ownership have drained.

The production predicate was completed on main by #37567 and #37883 while this PR was in draft. This PR now focuses on preserving that behavior with registered regression coverage and documenting the full idle contract.

Modifications

• Add registered scheduler tests covering all seven buffer-pipeline activity states:
  • pending backup batches
  • ongoing backup
  • ongoing write-through
  • pending load-back batches
  • ongoing load-back
  • ongoing buffer load-back
  • completed-ack queue
• Verify that each active state rejects flush without resetting the radix cache or KV allocator.
• Verify that flush succeeds once every pipeline state has drained.
• Preserve the scheduler health-check semantics while the pipeline is busy.
• Clarify the buffer-only idle contract in the scheduler comment.

No transfer, allocator, or cache implementation is changed.

Tests

Unit tests

PYTHONPATH=python .venv/bin/python -m pytest \
  test/registered/unit/managers/test_scheduler_buffer_mode_idle.py \
  test/registered/unit/managers/test_scheduler_flush_cache.py -q

Result:

10 passed, 7 subtests passed

The old predicate reproduces the regression for both missing states:

D2H: current_idle=False old_idle=True old_flush=True
H2D: current_idle=False old_idle=True old_flush=True

Formatting, Ruff, isort, codespell, registered-test validation, and git diff --check pass.

Ascend A3 lifecycle validation

Configuration: Qwen3-0.6B, Ascend backend, buffer-only HiCache, file storage.

Observed real transfer activity:

launched d2h node=10 tokens=4096
buffer busy ... d2h=1 storage=0
launched h2d ack_id=-3 tokens=4032
buffer busy ... h2d=1 ... d2h=0
prefetch fill committed ... filled=4032

Results:

first_cached_tokens:       0
flush_after_d2h:           HTTP 200, 3.258455s
replay_cached_tokens:      4032
flush_after_h2d:           HTTP 200, 13.0699s
post_flush_cached_tokens:  4032

Both flush calls waited for the corresponding pipeline state to drain before succeeding. Temporary tracing was removed after validation.

Checklist

• Rebased onto current main
• Registered focused regression tests
• Supported local test execution
• Real A3 D2H/H2D lifecycle validation
• No public API or documentation change