Shared prompt unification for Codex + Python agent

What changed

• Added prompts/lighthouse-shared.md as a single source of truth for the Lighthouse working rhythm, tools, and expectations.
• Updated scripts/codex-agent.sh to read the shared prompt file and inject it into the Codex CLI default prompt.
• Updated agent/lighthouse_agent.py to load the same shared prompt content into the system prompt used for Azure OpenAI / Anthropic runs.

Why it matters

• Keeps Codex CLI and the Python agent aligned on the same "Build → Reflect → Compact → Build" rhythm and concrete working instructions.
• Makes it easier to evolve our cultural norms and tooling guidance in one place instead of hand-editing multiple prompts.
• The change is fully covered by the existing test suite (./scripts/test-all.sh passes), so the agent behavior remains stable while getting clearer shared guidance.

Next ideas

• Add a very small note in README.md or HANDOFF.md pointing at prompts/lighthouse-shared.md so humans know where to tweak the shared instructions.
• Over time, consider factoring more of the agent-specific capabilities text (actions list, memory usage examples) into reusable prompt fragments if drift becomes a problem.