fix: persist repeated self-learning preferences

[EvanProgramming]

Summary

• Count each detected user-turn preference as independent evidence.
• Persist promoted preferences so a fresh process includes them in its prompt.
• Add a two-process regression test and update the self-evolution test-count snapshot.

Fixes #162

Validation

• Focused learning, worker, and evolution suites: 47 tests passed with the supported Python 3.12 disposable environment; the benchmark target also passed separately.
• Real CLI flow: main.py --project with /quit left the detached worker running; after a stale heartbeat it persisted learning.worker_started, learning.feature_started, learning.memory_scored, and learning.feature_completed with changed=true in SQLite.
• Real cross-process preference flow: two repeated observations promoted the proposal to active, and a fresh process emitted the learned language, naming-style, and verbosity preferences in its prompt.
• make check, make lint, make docs-check, make shell-check, and git diff --check passed.