fix(mcp): handle non-object checkpoint acknowledgements

[fallenmi]

What does this PR do?

Fixes #2272.

json.loads() accepts scalar and array roots, but tool_memories_filed_away() assumed the decoded checkpoint was an object. It unlinked the one-shot last_checkpoint marker and then called .get(), so MCP clients received an internal error with no acknowledgement marker left to retry.

This change:

• requires a decoded checkpoint to be a dict before reading msgs or ts;
• deliberately keeps unlink-before-validation, preserving the tool's existing consume-on-read contract;
• maps null, array, string, number, and boolean roots to the same content-free status: error response already used for malformed JSON syntax;
• preserves valid-object and missing-file behavior.

The scope is intentionally limited to valid non-object JSON roots. Invalid UTF-8, unlink races/failures, and malformed fields inside an object remain separate concerns.

How to test

• python -m pytest -q tests/test_mcp_server.py tests/test_mcp_http_transport.py tests/test_hooks_cli.py — 548 passed, 1 skipped
• Full split suite — 4,306 passed, 31 skipped
  • main suite: 4,297 passed, 31 skipped, 107 deselected
  • isolated cwd/install-sensitive subprocess cases: 9 passed
• pre-commit run --all-files

The regression matrix covers all five non-object JSON root types, a valid checkpoint object, invalid JSON syntax, a missing marker, marker consumption, and rejected-payload non-reflection. The existing HTTP read-only acknowledgement contract also remains green.

Checklist

• Tests pass
• No hardcoded paths
• Linter passes