[Store] Centralize client session lifecycle and resource ownership

[Aionw]

Description

Centralize client registration, remount readiness, heartbeat/liveness monitoring, host hints, and offboarding lifecycle in ClientSessionManager instead of MasterService.

• Add lock-scoped session admission, registration rollback, staged restore, and ordered state-change listeners.
• Keep resource cleanup in the MasterService listener while the manager owns monitoring, offboarding workers, and the snapshot cleanup barrier.
• Bind memory buffers and local-disk replicas to the exact client incarnation via read-only ClientSessionSharedPtr; replacing a session invalidates old segment generations rather than silently transferring their resources.
• Check local-disk staleness using the bound session rather than a point-in-time registry complement.
• Adapt snapshot restore, metrics, and tests. No RPC or snapshot wire-format change is intended.

Related work and scope

Searched existing issues/open PRs for client sessions and liveness before opening:

• [Store] Cut MasterService over to SegmentPool #3707 / RFC [RFC]: Refactor Mooncake Store segment management around SegmentPool #3360 reorganize mounted-region ownership around SegmentPool and explicitly exclude ClientSession redesign. This PR retains the existing segment manager and focuses on session ownership; the two changes will need integration coordination.
• [Store] Defer a client mass expiry the master cannot rule out as its own stall #3936 proposes mass-expiry deferral under master stalls. This PR does not add that policy.
• [Bugfix] Register SSD replicas per owner after client restart #4130 fixes SSD registration deduplication across different client UUIDs. This PR instead preserves exact session identity and centralizes lifecycle/admission.

Draft blockers

This is a substantive change exceeding 500 LOC. A dedicated RFC has not yet been filed/linked. Human confirmation of reviewing every changed line is also pending; keep this PR in draft until both are addressed.

Module

• Transfer Engine (mooncake-transfer-engine)
• Mooncake Store (mooncake-store)
• Mooncake Conductor (mooncake-conductor)
• Reshard (mooncake-reshard)
• Mooncake EP (mooncake-ep)
• Mooncake PG (mooncake-pg)
• Integration (mooncake-integration)
• P2P Store (mooncake-p2p-store)
• Python Wheel (mooncake-wheel)
• Common (mooncake-common)
• Mooncake RL (mooncake-rl)
• CI/CD
• Docs
• Other

Type of Change

• Bug fix
• New feature
• Refactor
• Breaking change
• Documentation update
• Performance improvement
• Other

How Has This Been Tested?

Local selected-target build and tests passed. HA/snapshot tests were adapted but not run in this submission; full integration and hardware-specific configurations remain unvalidated.

Test commands:

CCACHE_DIR="$HOME/.cache/ccache/mooncake" \
CCACHE_BASEDIR="$(git rev-parse --show-toplevel)" \
cmake --build build --target client_session_manager_test client_liveness_test \
  segment_test master_service_test promotion_on_hit_test -j 8
ctest --test-dir build --output-on-failure -j 4 \
  -R '^(client_session_manager_test|client_liveness_test|segment_test|master_service_test|promotion_on_hit_test)$'
prek run --files $(git diff --name-only HEAD^ HEAD)
git diff --check HEAD^ HEAD

Test results:

• Selected unit tests pass: 5/5 CTest targets (121.56 seconds).
• Integration tests pass (if applicable)
• Manual testing done (describe below)

All applicable lightweight hooks passed except mooncake-code-format: it requires unavailable clang-format 20 (the host provides 22.1.8). Changed-line clang-format 22 was used for the final inline-review edits, not claimed as a full PR formatting pass. The build emitted warnings from bundled fmt headers.

Checklist

• I have performed a self-review of my own code
• I have formatted my code using ./scripts/code_format.sh
• I have run pre-commit on the files changed in this PR and all hooks pass
• I have updated the documentation (if applicable)
• I have added tests to prove my changes are effective
• For changes >500 LOC: I have filed an RFC issue

AI Assistance Disclosure

• No AI tools were used
• AI tools were used (specify below)

OpenAI Codex assisted with code changes, inline-review follow-up, validation, and PR preparation. Human end-to-end review and responsibility confirmation remain pending; no claim of completed human review is made.