feat: add durable memory and natural preference learning

[Mouhand-Kaddo]

Summary

• Add durable, source-linked facts with revisions, recall, correction, and safe forgetting.
• Share project memory across worktrees while keeping session state isolated.
• Make compaction incremental and refresh memory in the active context.
• Learn preferences from natural messages without requiring a special phrase or standalone message.
• Add extraction diagnostics and /memory facts autocomplete.

Validation

• 1,529 tests passed locally.
• Ruff lint, formatting, and whitespace checks passed.
• Live testing confirmed that a preference within a multiline message is stored as a valid source-linked fact.

Behavior

• Automatic learning is opt-in with [memory] auto_learn = true and runs after successful compaction.
• Learned preferences must quote the original user message exactly and retain source evidence.
• Conflicting candidates are surfaced for review rather than silently replacing facts.
• Forgetting removes managed facts without deleting original transcripts or manually authored notes.
• Preference classification remains model-assisted.