CineStack 0.6.6 — Sampling host/pinned lifecycle

Evidence and implementation
- Inspected official ComfyUI v0.40.0 dynamic model load, low-VRAM patch pinning,
  HostBuffer ownership, prefetch cleanup and execution finally paths.
- Confirmed CineStack omitted reset_cast_buffers inside the multi-segment node.
  Added host-owned prefetch/reset sequence after every internal sampler call,
  including exceptions, before dropping patchers and optional model unloading.
- Preserve weights pin caches, sampling arguments, outputs, disk/cache identities,
  independent seeds, strict refine-only, rolling exports and bidirectional links.
- Add bounded metadata-only host probes and sampler pre/post cleanup checkpoints.
  Distinguish Comfy registered pins, weight/patch HostBuffers, PyTorch host cache,
  glibc, RSS and cgroup accounting. Do not sum overlapping measurements.

Local validation
- 164 Python tests: 162 passed, 2 skipped (no system FFmpeg); bundled FFmpeg runs
  rolling/export tests. Full run: 37.662 seconds.
- Final cleanup-status labeling adjustment: all 9 host-memory tests passed again.
- 134 real Chrome assertions: 42 seed/version/migration, 82 Director controls,
  10 versioned asset/cache checks. Asset version v0.6.6-bb1de5cb0b6c187d.
- Release identity/example versions and CRLF-aware whitespace checks passed.
- Nine new tests cover real sampler dispatch against a CPU pin lifecycle fixture,
  19 internal stages, patch strong-owner release, base-weight pin preservation,
  unchanged seed/output behavior, sampler/cleanup failures, legacy API fallback,
  unload order, shared pin-state dedup, bounded probes, phase-specific snapshots,
  PyTorch host isolation and ContextVar restoration on success/failure.
- Existing official-node import AST test now supplies package context for its
  new relative helper imports; its sampler/shift dispatch assertions are unchanged.
- Existing CPU tests exercise rolling/partial/full exports, strict second-pass
  reuse/failure, independent seeds, connected/disconnected junctions and old plans.

Limits and acceptance
- Native CUDA registration, GPU model execution, HostBuffer physical RSS and
  performance are not reproduced on this Mac. Lifecycle fixtures are not GPU proof.
- The old server trace did not measure pinned allocations. The missing cleanup is
  confirmed in code, but attributing every observed ~1.7 GiB to patch pins requires
  the next server run. Base-weight warmup/other native caches can also grow.
- See host-memory-lifecycle.txt for source links and precise 19-segment acceptance.
- No Registry publication; GitHub push only. CI result is available on the commit.
