Measured in the Codex desktop app: a fresh folder per conversation meant
an orphan ledger per conversation - records vanished when the chat closed.
The default now follows one rule: project evidence decides. Git repo or
existing .argus -> project ledger, unchanged; anywhere else -> the
personal home ledger (ARGUS_HOME or ~/.argus), announced once on stderr.
Verified live: a prediction saved in one simulated app conversation is
visible from a second; no orphan created. Explicit argus_dir/ARGUS_DIR
still win. Target model: one logical judgment dataset per user (ADR
2026-07-27) - this removes the sharpest fragmentation against it.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016piraK9GPUy7UYbHnwCPzD