[shipiyouniao]

Inherits #574's tests through #455; the tests are not duplicated in this PR's own commit.

After a confirmed physical-capacity rejection, repeating the same growth transaction adds IPC without creating memory. This optional layer over #455 delays repeated single-page growth and wakes it after a successful physical release changes a per-GPU token.

Mapped free blocks and resident reservations remain usable during backoff. Tokens are best-effort hints, not capacity grants; missing signals fall back to timer probes. Notifications are published only after ordered unmap succeeds. The native ledger remains the authority for each new grant.

Rebased onto the refreshed #455, preserving main's lifecycle, resize/reclaim and counter implementation. A rejected multi-page background preallocation no longer suppresses smaller foreground growth: failure to reserve several pages does not prove one page is unavailable. This narrows the existing retry gate without adding an RPC or lock to successful allocation.

Dependency: #455; #489 is an independent sibling. Review this PR's own change against 2cc3240, rather than its cumulative main diff.

Private validation at the final head: Python 3.11 CPU 1,553 passed / 15 skipped; rebuilt native extension on one GPU 98 passed / 3 two-GPU tests skipped. C++ checks, all-file pre-commit and strict MyPy targets 3.10-3.13 passed. Dedicated Codex review found no actionable regressions; its local platform limitations are covered by the private Linux runs. The updated head will run the Python 3.10-3.13 PR CI matrix.

No new serving-throughput or multi-GPU claim. Draft state remains unchanged.