fix(delegation): bind subagent.interrupt to the caller's session, like subagent.steer

[pierrenode]

Summary

This week's 60e1f7517/a94ebf5f5/9d4ef04ed series hardened steer_subagent() to require exact owner_session_id + owner_transport (ContextVar-bound, unforgeable) + owner_session_record identity before a steer is accepted, closing a cross-session authority hole. interrupt_subagent() — which steer_subagent()'s own docstring says it mirrors ("The redirection-side mirror of interrupt_subagent()") — never got the same treatment: it takes only subagent_id and performs zero ownership validation. The subagent.interrupt RPC handler doesn't even resolve a caller identity before calling it.

_active_subagents is a bare module-level dict, not session- or profile-scoped. delegation.status (list_active_subagents()) returns every active subagent's id system-wide with no session filter, so any session can enumerate ids and then hard-stop a child it does not own via subagent.interrupt.

Fix

Mirrors steer_subagent()'s exact ownership contract:

• tools/delegate_tool.py::interrupt_subagent() gains owner_session_id/owner_transport/owner_session_record kwargs, checked identically to steer_subagent(). owner_session_id=None (the default) deliberately preserves the internal in-process helper contract — the one existing caller (tests/agent/test_interrupt_compat.py) needs no changes.
• tui_gateway/methods_session.py's subagent.interrupt handler now resolves caller authority via _current_session_steer_authority(), identical to subagent.steer's handler.
• The only production client, ui-tui's agentsOverlay.tsx, never sent a session_id at all — without threading it through, the ownership check would have silently broken the interrupt button for every legitimate caller. Added a sessionId prop (AgentsOverlayPane → ui.sid, mirroring ModelPicker's existing sessionId threading pattern) and pass session_id in the RPC params.

Testing

• Added registry-level ownership tests for interrupt_subagent() (owned succeeds, foreign session denied, foreign transport with correct session_id denied) and a TestSubagentInterruptRPC class mirroring the existing TestSubagentSteerRPC coverage to tests/tools/test_subagent_steer.py.
• Mutation-verified: stashed the backend fix, confirmed all 7 new authority-check tests fail on pre-fix code.
• tests/tools/test_subagent_steer.py (41), tests/agent/test_interrupt_compat.py, tests/tools/test_delegate.py + 5 neighboring delegate/subagent suites (96 tests), and the full tests/test_tui_gateway_server.py (517/518 — the one failure is a pre-existing threading-timing flake in an unrelated write_json test, confirmed by isolated re-runs and against pre-fix code) all pass.
• ui-tui: npm run typecheck clean; full vitest suite 138/138 files, 1528 passed.
• ruff check clean on all changed Python files.

Note on a related open PR

#70899 ("Mission Control" async-delegation panel, branched 2026-07-24 before the steer_subagent hardening landed) adds a parallel send_to_subagent() whose docstring explicitly states it matches interrupt_subagent()'s current no-ownership-check posture, and touches the same four files this PR touches. The maintainer has commented it will be salvaged/expanded. This fix doesn't depend on or block that PR — it hardens a real, currently-exploitable gap in main today — but whichever lands second will likely need a rebase against the other.