Apply SecretDetector to the extraction payload before LLM submission

[troylar]

Extraction (`extractor.py`) reads raw instruction-file contents and formats them directly into the LLM prompt. `SecretDetector` currently guards only MCP-config templating in `package_creator.py` — the extraction payload itself is never scanned.

Hardening ask:

• Run the extraction payload through secret detection (redact/template hits) before any provider call, on both the AI path and anything the non-AI path persists
• Add a test that asserts a planted secret in an instruction file does NOT appear in the outbound payload

Tracked as DEBT-001 in `.codehealth/REGISTER.md`.

[troylar]

Fix landed: redact_secrets_in_text added to secret_detector.py (assignment-pattern + provider-token scanning over free text, HIGH-confidence redaction only) and wired into PracticeExtractor._read_instruction_files — one choke point covering both the AI and file-copy paths. Two tripwire tests plant secrets and assert they never reach the LLM payload or persisted package, plus an anti-over-scrubbing assertion. Full suite: 1684 passed.

Register status: resolved-pending-verify (closes after the next passing measurement, per .codehealth soak rule).