fix: harden recovery and storage behavior

[davidrobertson]

Summary

• Keep post-compaction recovery pending until lcm_pack_context completes.
• Redact credential URI passwords, including password-only Redis URIs, before JSONL and SQLite persistence.
• Restrict LCM storage and recovery state to 0700 directories and 0600 files.
• Make fresh-store cleanup a successful no-op and search each integrity-checked overflow payload through its existing 8 MiB safety bound.
• Preserve the retrieval-quality benchmark floor and document the full overflow scan.

Root causes

The recovery marker was consumed when the reminder was shown, so a task could stop before retrieval completed. URI redaction assumed a nonempty username, storage permissions depended on the caller's umask, cleanup expected an existing index, and overflow search reused the 512 KiB paging limit intended for display.

Validation

• npm test: 188 passed
• npm run typecheck: passed
• npm run smoke: passed
• Direct CLI/MCP audit: passed under umask 000
• Goal, QA, code-quality, context/integration, and security review lanes: passed at 3559b2e897f64e4e41141c93c4e4f124cd88ca82

Risk

Overflow search now reads and lowercases each managed payload through the existing 8 MiB input bound, so broad searches across many large payloads can use more CPU. Recovery marker writes still trust the private LCM home; a hostile preexisting symlink requires prior control of that directory.