Add deterministic redaction and reduce Auto Memory logging

Problem

Auto Memory reads local transcripts and sends selected transcript content to the model used by the background extraction agent. The extraction prompt instructs the model to redact secrets, but that happens after content is already in model context. The service can also log existing skills and pending inbox content.

Relevant code:

• packages/core/src/agents/skill-extraction-agent.ts
• packages/core/src/services/memoryService.ts

Impact

• Secrets or sensitive local details present in transcripts may be sent to the model before deterministic redaction happens.
• Debug logs may contain memory content, pending patch contents, absolute paths, or sensitive workflow details.
• The current user-facing setting description does not clearly state that Auto Memory performs model calls over local transcript content.

Expected Behavior

Auto Memory should scrub sensitive content before model submission and avoid logging full memory or patch content by default.

Acceptance Criteria

• Add deterministic pre-model secret scrubbing for transcript excerpts, session summaries, workflow hints, and pending inbox content.
• Avoid logging full memory or patch content by default.
• Keep enough diagnostic information for debugging without dumping sensitive content.
• Update settings/docs to clearly explain that Auto Memory uses model calls over local transcript content.
• Tests cover common secret patterns in session summaries, workflow hints, and pending patch content.

Notes

This does not require changing the review-first inbox model. It tightens the privacy boundary before extraction prompts are assembled and before debug logs are emitted.