Filter negated evidence from Aragorn creative results

[cbizon]

Summary

• Filter Aragorn creative-mode bindings when the bound edge or any recursively nested support edge has biolink:negated=true.
• Exit after one cheap knowledge-edge scan when the message contains no negated assertions; no results or support graphs are traversed in the common case.
• For affected messages, build reverse edge and auxiliary-graph dependencies once, propagate negated and incomplete-support status, and check each binding with constant-time set membership.
• Preserve an analysis when a query edge still has another valid parallel binding; remove the analysis only when a required query edge has no valid bindings left.
• Scope the behavior to Aragorn single-edge inferred queries so other Shepherd targets are unchanged.

Testing

• uv run --no-sync pytest tests/unit/test_merge_message_helpers.py tests/unit/test_merge_message.py tests/unit/test_merge_message_batch.py -q (45 passed).
• uv run --no-sync pytest tests/unit --ignore=tests/unit/test_cpu.py -q (447 passed).
• uv run --no-sync black --check workers/merge_message/worker.py tests/unit/test_merge_message_helpers.py.
• git diff --check.

[cbizon]

I worry that this approach will be slow for large queries. @maximusunc what do you think?

[cbizon]

Follow-up optimization in b259d9f:

• The common no-negation path now scans knowledge-graph edges once and exits before traversing any results or support graphs.
• Messages containing negated assertions build one reverse edge/auxiliary-graph dependency index and propagate affected support chains once.
• Result filtering then uses constant-time set membership instead of recursively walking each binding.
• Parallel bindings remain independent: a negated binding is removed while a valid binding for the same query edge keeps the analysis alive.
• Verification: 45 merge tests and 447 applicable unit tests pass.