[security] fix(personalization): stop replaying exported env var values

[shaun0927]

Closes #149.

Summary

This PR narrows personalization env-var extraction so exported environment variables are stored by name only, not by raw NAME=value payload.

Concretely:

• export OPENAI_API_KEY=... is now remembered as OPENAI_API_KEY
• raw env-var values are no longer persisted into local_rules
• later runtime system prompts no longer replay those raw secret values
• regression tests cover both extraction and prompt assembly

Why

PR #65 introduced personalization so OpenHarness can remember local environment context such as hosts, paths, and endpoints. That feature direction makes sense.

The problem is that the current implementation also captures full exported values. For secrets, that creates a different class of behavior:

• the value is written to facts.json
• the value is rendered into rules.md
• the value is replayed into future build_runtime_system_prompt(...) output

So a one-off shell snippet like export OPENAI_API_KEY=... can become a persistent prompt-side secret disclosure path.

Root cause

src/openharness/personalization/extractor.py currently captures the full NAME=value payload for export ... patterns.

That makes sense for general parsing, but it is too broad for a personalization feature whose output is persisted and later injected into prompts.

Change

• update env-var extraction to keep only the variable name
• keep the rest of the personalization flow unchanged
• add a regression test showing the extractor does not retain secret values
• add a prompt-level regression test showing exported secret values are not replayed into the runtime system prompt

Before / After

Before

• export OPENAI_API_KEY=sk-test-secret -> personalization stores OPENAI_API_KEY=sk-test-secret
• later sessions replay sk-test-secret via local rules in the system prompt

After

• export OPENAI_API_KEY=sk-test-secret -> personalization stores OPENAI_API_KEY
• later sessions can still see that the variable exists, but not the secret value

Validation

• PYTHONPATH=src pytest -q tests/test_personalization/test_extractor.py tests/test_prompts/test_claudemd.py
• PYTHONPATH=src ruff check src tests
• targeted regression coverage added for extraction + prompt assembly

Notes

• I intentionally kept this PR narrow: it fixes the confirmed secret-persistence path without redesigning the whole personalization feature.
• Full pytest -q on this local environment is currently blocked by unrelated collection errors caused by a missing optional pyperclip dependency; I did not change any of the affected command/UI code paths.