Stop worker-owning nodes leaking forever when New Chat interrupts a generation

[dovvnloading]

Summary

Started as a narrow fix (give ArtifactNode a GC-time __del__, matching CodeSandboxNode) and grew once verification showed the __del__ alone wouldn't do anything.

Root cause 1: ChatScene._teardown_items_before_clear() already calls dispose() explicitly for chart_nodes on every clear() (New Chat / chat-switch), but never did for any worker-owning node type (artifact/conversation/pycoder/code_sandbox/gitlink) — it only stopped their hover-animation timers. A GC-time __del__ backstop is inherently non-deterministic, and (per root cause 2) sometimes can't fire at all. _teardown_items_before_clear() now calls dispose() on all five worker-owning node lists, mirroring the existing chart_nodes treatment — New Chat now stops an in-flight generation the instant it happens.

Root cause 2 (found while checking whether the __del__ fix would even matter): ArtifactNode was not collectible by Python's GC at all, ever. Two independent causes, both confirmed empirically:

• ArtifactNode's own instruction_input.submit_requested was wired via a lambda closing over self on a custom Signal — fixed with a bound method.
• graphlink_window_actions.py wires each of ArtifactNode's/ConversationNode's/PyCoderNode's/CodeSandboxNode's worker's own signals via a lambda carrying node=/thread= as default args. PySide6's GC does not reclaim a lambda connected to a custom Signal (a bound-method connection to the same signal is reclaimed fine) — so as long as those connections stood, both the worker and the node were immortal for the rest of the process, regardless of dispose(). GitlinkNode was the one sibling whose dispose() already disconnected the worker's own signals; that pattern is now applied to the other four.

Changes

• dispose() on ArtifactNode/ConversationNode/PyCoderNode/CodeSandboxNode now disconnects the worker's own signals (mirroring GitlinkNode) before nulling the reference
• ArtifactNode's internal lambda replaced with a bound method
• ArtifactNode/ConversationNode gained a guarded __del__ (defense-in-depth, matching CodeSandboxNode's existing convention — no longer the primary mechanism now that clear() calls dispose() deterministically)
• ChatScene._teardown_items_before_clear() calls dispose() on all 5 worker-owning node lists

Test plan

• Adversarial verification workflow (2 independent agents, 120 tool calls) re-confirmed the fix and reproduced every leak empirically with the real worker classes — no mocks, since a mock silently absorbs a bad connect() call and hides this exact bug class
• New regression tests reproduce the real graphlink_window_actions.py wiring for all four node types and prove both the node and the worker become GC-collectible after dispose()
• Full suite: 1705 passed, zero regressions
• Real app launch verified clean (zero errors in the log)

The workflow also found the identical lambda-over-self pattern in three unrelated context-menu classes and ColorPickerDialog — a different feature area (right-click menus, not worker-owning nodes), not touched here; flagged separately as a follow-up task.

Third of the fixes from the recent adversarial bug-scan (after #91, #92).