feat: strengthen retrieval eval and runtime integrations

[cafitac]

Summary

• expand retrieval evaluation into a stronger regression/comparator harness with symbolic fixture references, comparator matrix baselines, selective baseline gating, soft advisories, richer task metadata, and summary/delta rollups
• add Codex and Claude prompt attach surfaces plus reusable wrapper scripts, and document real runtime readiness status
• fix Hermes hook bootstrap merge behavior so existing hook indentation and continuation lines are preserved for both fresh-install and existing-hooks users

Key changes

• retrieval-eval
  • support symbolic top-level fixture references plus richer selectors
  • add baseline modes: lexical, lexical-global, source-lexical, source-global
  • add selective --fail-on-baseline-regression-memory-type gating
  • add advisory-only soft thresholds for regression and baseline regression
  • expose rationale, notes, by_memory_type, by_primary_task_type, top-level pass/fail counts, and richer delta rollups
  • add adversarial checked-in fixtures for stale/drift/guardrail scenarios across facts, procedures, and episodes
• runtime attach surfaces
  • add codex-prompt and claude-prompt CLI commands
  • add scripts/run_codex_with_memory.py and scripts/run_claude_with_memory.py
  • verify Codex wrapper E2E; keep Claude Code marked as pending auth-gated E2E
• Hermes integration
  • preserve event-local indentation when merging hook snippets into existing Hermes configs
  • document fresh-install, existing-hook merge, doctor/test flow, and first-run --accept-hooks verification

Verification

• uv run pytest tests/test_prompt_wrapper_scripts.py -q
• uv run pytest tests/test_cli.py -q
• uv run pytest tests/test_cli.py tests/test_retrieval_evaluation.py -q
• uv run pytest tests/ -q
• real Hermes verification:
  • uv run agent-memory hermes-bootstrap ~/.agent-memory/memory.db --config-path ~/.hermes/config.yaml
  • hermes hooks doctor
  • hermes hooks test pre_llm_call
  • hermes --accept-hooks chat -q 'Reply with OK only.' --quiet
• real Codex verification:
  • codex exec --skip-git-repo-check --sandbox workspace-write --model gpt-5.4-mini 'Reply with OK only.'

Readiness status

• Hermes: ready for real use
• Codex: ready for real use
• Claude Code: prompt/wrapper surface present, but final sign-off still blocked on auth-enabled E2E