feat: update hermes to 1.5.3; add dataset-switch, forget, recall scopes, code graph

[siillee]

hermes-agent 1.2.0 — parity with the claude-code/codex/openclaw plugins

Brings the Hermes plugin up to the other cognee agent plugins on every
user-facing memory operation:

• Layered recall: per-prompt fan-out over session cache, trace lessons,
  agent guidance and graph (was one graph-only auto call), with a budget,
  labelled blocks, and a plain-words memory-hit header.
• Per-document forget: two-phase find → confirm → delete; dataset-wide
  deletion gated, all-datasets wipe not expressible from the tool.
• cognee_switch_dataset: move a conversation to another dataset;
  bridges the old session, persists the override across resumes.
• Code graph: hermes cognee index-repo CLI, cognee_code_search tool,
  identifier-gated code recall lane (explicit indexing only, like openclaw).
• Memory steer, and a version command + rate-limited PyPI update check.
• cognee pinned to exactly 1.5.3 (==), required for code indexing and
  targeted forget invalidation; check_version_pins.py now accepts exact
  pins (repo-level tooling change, no effect on other integrations).

[github-actions]

A 2,964-line parity release with exactly one dataset — impressive restraint, considering how much you could have broken. 📚

🟢 No blocking issues found.

What this PR does well:

• Security: UUID validation prevents injection; two-phase forget with confirm=true prevents accidental deletion
• Data safety: Forced dataset switches defer failed bridges to session end instead of dropping turns; cognee session never spans two datasets
• Test coverage: 87 test classes across 17 files, including dedicated tests for layered recall, dataset switching, code graph, and update checks
• Breaking change handling: cognee pinned to exactly 1.5.3 is well-justified (code graph + targeted session invalidation require it) and clearly documented in CHANGELOG
• Fail-safe design: Update check is network-optional and cached; dataset override persistence is best-effort

Findings:

• minor — integrations/hermes-agent/cognee_integration_hermes/dataset_overrides.py:52 — File write has theoretical race condition in multi-process scenario; mitigated by best-effort design and single-process usage pattern

See inline comments for details.

---

Fix the review findings in PR #385:
1. [minor] integrations/hermes-agent/cognee_integration_hermes/dataset_overrides.py:52 — (Optional) Use atomic write via temp file + rename to make dataset override persistence safe against concurrent writes