fix(graph): scope tunnel reads and deletes to config

[fallenmi]

What does this PR do?

• threads the optional palace config through tunnel list, follow, and delete paths
• keeps deletion's lock, load, and save on the same selected tunnels.json
• passes the MCP server's canonical _config through tunnel and adjacent hallway handlers
• adds same-ID regressions across two palaces, proving reads and deletion cannot cross palace boundaries
• documents the fix in the changelog

The root cause was an incomplete config-propagation path: create_tunnel(..., config=selected) already wrote to the selected sidecar, but matching readers either ignored or could not accept that config. The hallway MCP handlers had the same adjacent gap. With ambient palace B active and MCP palace A selected, reads could return B's same-ID record and deletion could mutate B while leaving A untouched.

Calls that omit config retain the existing default-palace behavior. This completes the tunnel read/delete side of the palace-isolation work merged in #2018.

Closes #2263

Reproduction and verification

Exact current develop base a9f345cc63254eb4dea7abad36963b85c9f8453a:

• the two-palace hallway oracle is RED: MCP reads B and deletes B while selected palace A is unchanged

Exact head f16a7dfc04f55bf764fcdd1b8bda7cd3cbdd3f1e:

• the same oracle is GREEN: MCP reads and deletes only A; B's same-ID record remains intact
• adjacent production-consumer audit confirms hallway config forwarding throughout
• relevant suite: 478 passed
• focused post-commit rerun: 5 passed
• full suite: 4,591 passed, 32 skipped, with one ambient peer-writer-state failure that passed immediately in isolation
• Ruff check and format, pre-commit, and git diff --check passed

AI assistance

OpenAI Codex was used to rebase the existing contribution onto current develop, implement the maintainer-requested adjacent hallway fix, run the exact-base RED / exact-head GREEN checks, and execute the test and lint commands recorded above. Igor reviewed the published scope and evidence.

Checklist

• Tests pass
• No hardcoded paths
• Linter passes

[igorls]

This tunnel isolation fix is important, but it leaves the adjacent hallway MCP handlers with the same cross-palace mutation bug.

On current develop, list_hallways and delete_hallway already accept an optional config, but tool_list_hallways calls list_hallways(wing) and tool_delete_hallway calls delete_hallway(hallway_id) without forwarding the server _config. I reproduced this with MCP _config selecting palace A and the ambient MempalaceConfig() selecting palace B: both sidecars contained the same hallway ID; tool_delete_hallway returned deleted: true, left A unchanged, and deleted B.

Please extend this contribution to pass config=_config through both hallway handlers and add the same two-palace read/delete regression coverage used for tunnels. The branch also needs rebasing; the current merge conflict is in CHANGELOG.md.

[fallenmi]

The requested changes are now on exact head f16a7dfc04f55bf764fcdd1b8bda7cd3cbdd3f1e: rebased onto current develop a9f345cc63254eb4dea7abad36963b85c9f8453a, both hallway MCP handlers forward the server _config, and same-ID two-palace read/delete regressions prove palace B remains untouched. The current-base oracle is RED and the head is GREEN; 478 relevant tests, Ruff/format, pre-commit, and diff checks pass. The full suite had 4,591 passed / 32 skipped plus one ambient peer-writer-state failure that passed immediately in isolation.

The PR body now contains the exact updated evidence. Disclosure: this follow-up and verification were prepared with OpenAI Codex assistance and reviewed before publication.