v0.6.3

AID Interpret step is now strictly private.

The AID cycle was designed with Attend and Interpret as private calibration steps — the agent reads engagement signals, forms a hypothesis, adjusts its response. Skill text was narrating interpretations back to the user ("You asked about X", "You engaged with Y") as preambles to probes. This is content-level sycophancy dressed as observation — functionally identical to "Great question!" — and corrupts the AID architecture by making private calibration public.

Probes and challenges now reference specific artifact content directly ("The essay's conclusion on X rests on Y", "ADR-NNN trades A against B") rather than characterizing the user's prior engagement. Updated across the orchestrator and all six phase skills.

The third consecutive patch release from practice — sycophancy hides in structure, not just tone. Each of these was a place where the skill text said the right thing in principle but left operational slack. Cutting the slack turns principle into forcing function.