cortexkit-lease: shared-mode leases (acquire_shared)

[ualtinok]

What

Adds LeaseStore::acquire_shared: shared holders coexist with each other and block the exclusive writer (and vice versa). File impl via fs2's try_lock_shared — flock(LOCK_SH) on unix, LockFileEx without LOCKFILE_EXCLUSIVE_LOCK on Windows.

Why

First consumer: Synapse's cross-module model cache (~/.local/share/cortexkit/models/). Readers take a shared lease on a blob digest while validating/mmap-ing; GC takes exclusive + two-phase tombstones, so it can never delete a file under a live reader while concurrent readers never serialize each other. (Design: synapse docs/design-synapse-module.md, model-cache section; requested shape per SUBC review pm_31c3ffbf — lease-crate-only, option (a).)

Semantics pinned by tests

• shared + shared coexist; shared blocks exclusive until the last shared holder drops; exclusive blocks shared
• shared handles do not bump the fence epoch (not writers) — proven epoch-neutral; they report the last persisted writer epoch for observability
• cross-process check (unix): a python-fcntl child holds LOCK_SH while the parent's exclusive acquire is Held and a second shared coexists — pins real OS semantics, not per-fd artifacts. On Windows the same-process tests exercise LockFileEx contention directly (its locks are per-handle).

Notes

try_lock_shared is called fully-qualified through the fs2 trait: std 1.89+ ships an inherent File::try_lock_shared returning TryLockError which otherwise shadows the fs2 method (build error caught in this PR's development.

No changes outside the lease crate. cortexkit-store untouched and green.
EOF
)

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}

---

Summary by cubic

Add shared-mode leases to cortexkit-lease via LeaseStore::acquire_shared, letting multiple readers coexist while blocking an exclusive writer (and vice versa). Prevents model cache GC from deleting files under live readers without forcing readers to serialize.

• New Features
  • acquire_shared: shared+shared coexist; exclusive blocks shared; exclusive waits until the last shared drops.
  • Epochs: shared handles don’t bump the write fence; LeaseHandle::epoch() reports the last persisted writer epoch.
  • File backend: uses fs2 try_lock_shared (flock LOCK_SH on Unix, shared LockFileEx on Windows); called fully qualified to avoid std File::try_lock_shared shadowing.
  • Tests: coexistence/blocking, epoch neutrality, and a Unix cross-process check; Windows semantics covered via per-handle locks.

^{Written for commit dda5aee. Summary will update on new commits.}