Scope API sessions by session key

[necoweb3]

This records X-Hermes-Session-Key ownership for API-server session resources and enforces that owner key across the /api/sessions resource surface.

Why

X-Hermes-Session-Key is the API server's stable per-user/channel boundary for shared-key frontend deployments. The chat endpoints parsed and forwarded it to the agent, but the session resource endpoints selected sessions by session_id only. A caller with the same bearer token and a different session key could list, read, patch, delete, fork, or continue another keyed session if it knew the session_id.

Wrong-key requests now return 404 so the session's existence is not disclosed.

Changes

• Store gateway_session_key in session model_config when creating keyed API sessions.
• Filter GET /api/sessions by the caller's session key.
• Enforce owner checks on:
  • GET /api/sessions/{session_id}
  • PATCH /api/sessions/{session_id}
  • DELETE /api/sessions/{session_id}
  • GET /api/sessions/{session_id}/messages
  • POST /api/sessions/{session_id}/fork
  • POST /api/sessions/{session_id}/chat
  • POST /api/sessions/{session_id}/chat/stream
• Preserve legacy ownerless sessions while claiming ownerless sessions on first keyed chat use.
• Add regression coverage for wrong-key list/get/messages/patch/fork/chat/stream/delete attempts.

Tests

python -m pytest tests/gateway/test_session_api.py -q --timeout-method=thread
12 passed

[liuhao1024]

LGTM. The session-scoping logic is well-structured:

• _session_visible_to_key correctly allows unscoped sessions (no gateway_session_key) to remain visible to all callers for backward compatibility.
• _claim_session_owner_if_needed lazily assigns ownership on first keyed access, so existing sessions aren't broken.
• The list endpoint filters post-query, which is simple and correct for the current data volume.
• _parse_session_key_header reuse keeps the header parsing consistent across endpoints.

One minor note: _session_visible_to_key compares owner_key == gateway_session_key as strings. If the header value arrives with leading/trailing whitespace, the equality check will fail silently and the caller will get a 404. Consider .strip() on the parsed key. Not blocking — the header parser likely already normalizes this.

[necoweb3]

Rebased this PR onto current `origin/main` and reran the focused session API coverage.

Validation:

```text
python -m pytest tests/gateway/test_session_api.py -q --timeout-method=thread
13 passed