Add attention sovereignty and input immunity scoring

[sparklkt]

What changed

This PR adds attention sovereignty and input immunity scoring ahead of disturbance evaluation.

• adds AttentionAssessment and a rule-based attention assessor
• scores manipulative signals, low-value stimulus pressure, novelty traps, emotional hijack risk, goal relevance, long-horizon value, and safety criticality
• adds recommended_depth to guide whether an input deserves minimal, standard, or deep processing
• feeds attention assessment into disturbance scoring so low-value high-stimulation inputs are downweighted while safety-critical inputs are not ignored
• records attention assessment in growth logs and exposes it through the pipeline result and CLI output
• adds tests for low-value downweighting, safety-critical preservation, long-horizon prioritization, and growth-log persistence

Why it changed

The prototype needed a way to protect attention from novelty traps, manipulative prompts, and low-value stimulation without suppressing real safety issues or long-horizon research tasks. This PR adds that gate before the rest of the reasoning pipeline.

Validation

• .venv/bin/pytest -> 20 passed in 0.12s

Safety boundary

• remains offline and rule-based
• no LLM integration
• no network participation in reasoning
• no real external actions
• no bypass of human judgment
• safety-critical inputs still surface strongly even when emotionally charged