feat: add feedback evaluation and memory safeguards

[jayzeng]

Summary

• add an external-feedback evaluation dataset and deterministic capability probes, with optional isolated multilingual qmd retrieval probes
• harden persisted and injected memory with provenance, secret redaction, lifecycle/trust filtering, and a strict 16,000-character context cap
• add opt-in prompt-aware context retrieval and ship a portable Node.js CLI for global npm installs
• document product boundaries, evaluation usage, and the updated memory/context behavior

Tests

• bun test test/unit.test.ts
• bun test test/cli.test.ts
• bun test test/eval.test.ts
• bun run lint
• bun run build

Storage and qmd behavior

• persisted entries may now include an optional standalone Source: line when --source-uri is supplied
• recognized secret shapes are redacted before persistence and context output
• inactive, expired, revoked, retired, superseded, or explicitly untrusted entries remain on disk but are excluded from context, distillation, and retrieved qmd snippets
• context --query opts into qmd-backed retrieval; timed-out searches are aborted and unresolved or policy-invalid source paths fail closed