CineStack 0.6.4 — Stage-level memory diagnostics

Changes
- Read-only observations enabled by default for rolling generation; env opt-out.
- Console summaries and per-run memory_diagnostics.jsonl, including partial/error traces.
- Process RSS/anonymous/file/swap, available glibc allocator counters, cgroup usage,
  CUDA allocator counters without initialization/synchronization, model record counts.
- Bounded explicit tensor traversal with backing-storage deduplication, slices,
  Comfy AV wrappers, cycles, category sizes, largest storages and truncation markers.
- Hooks around conditioning, sampling/cache reuse, refine passes, decoding, encoding,
  old-reference cleanup and final merge. Sampling and cleanup policies unchanged.

Verification
- Release identity and example version validation passed.
- Diff whitespace checked with cr-at-eol for existing CRLF source files.
- Python full suite: 148 tests; local environment skips two system-FFmpeg audio tests.
- Browser: 42 seed/version/migration + 82 workbench + 10 versioned-assets checks passed.
- Diagnostic cases cover shared storage, sliced views, distinct clones, AV wrappers,
  bounded traversal, weak-reference lifetime, exact phase deltas, no prompt contents,
  disabled behavior, unavailable probes, write errors, execution-stage ordering,
  bounded rolling frames, seed/cache identity, strict refinement reuse, real N-pass
  refine-loop observation without enabling intermediate export, and failure traces.

Limits
- Host/model GPU inference is simulated in CPU executor tests.
- Existing continuity/rolling/partial-export regressions remain included.
- A partial storage inventory is not the whole Python/native heap or resident RSS.
- Runtime peaks between observation points can be missed; external/native allocators
  may be outside mallinfo2 coverage. No malloc_trim or added model unloading.
- GPU reproduction and root-cause confirmation require the user's updated server run.
- No Comfy Registry publication invoked.
