feat(sse): persist bounded resume cursors

[luoye520ww]

Problem

Renderer SSE acknowledgement state is currently in-memory. A renderer/runtime restart loses the last confirmed sequence and can replay an unnecessarily large thread window.

Root cause

There is no bounded, atomic persistence layer for per-scope resume cursors.

Scope

This PR adds only the shared cursor contract and a main-process persistence service. It is intentionally independent from IPC and renderer recovery wiring.

Changes

• Add strict, versioned SseResumeCursor and file envelope schemas.
• Persist at most 256 scopes with atomic writes and serialized mutations.
• Reject same-generation cursor regressions.
• Allow a new runtime generation to reset its sequence.
• Keep malformed files untouched and fail open to authoritative replay.
• Ignore storage entries whose key does not match their scope.

Non-goals

• No preload/IPC bridge changes.
• No renderer subscription changes.
• No automatic cursor restore yet; that is the follow-up integration PR.

Safety

• Cursor data contains only bounded IDs, sequence numbers, generation metadata, and timestamps.
• Writes use the existing atomic file helper.
• Corrupt data is never overwritten automatically.
• Entries are capped to prevent unbounded user-data growth.

Typecheck

npm.cmd run typecheck
npm.cmd --prefix kun run typecheck

Both passed.

Tests

npm.cmd exec vitest run src/main/services/sse-resume-cursor-store.test.ts

Result: 1 file, 5 tests passed.

Also passed:

npm.cmd run lint
npm.cmd run build
git diff --check

Review performed

• Functional correctness
• Concurrency and lifecycle
• Data integrity and atomic writes
• Corrupt-file and generation-boundary handling
• Cross-platform path handling
• PR scope and diff cleanliness

Dependency

Depends on #911 (and its sequence/gap contracts in #902/#910).

Issue

Part of #885

[luoye520ww]

Closing this stacked version in favor of a clean develop-based PR; the cursor store does not depend on the earlier SSE implementation commits.