agentloop init separates durable evidence from per-machine state
agentloop init now writes a .agentloop/.gitignore that keeps per-machine harness state out of git — state.json (the active-task pointer), loops/<instance>/ loop directories, and the local runs/ ledger — while durable evidence (task contracts, reports, handoffs, policies, loops/*.md templates, manifest.json) stays tracked. This stops the harness's own writes from dirtying the working tree and being read as scope drift by tools that watch it. The .gitignore is scaffolded skip-if-exists like every other generated file; --json/--redact-paths behavior is preserved.
When re-running agentloop init over a repo that committed per-machine state under an older version, init now detects the already-tracked paths and prints the exact git rm -r --cached … command to un-track them (using a :(glob) pathspec so it targets loop instance directories without touching the tracked loops/*.md templates). Detection is read-only — init never runs git rm itself — and surfaces in both human output and --json (perMachineState).
Behavior note: because .agentloop/runs/ is now git-ignored, run-ledger files no longer appear as untracked working-tree changes in agentloop status / guard output. This is intended — local run evidence is no longer treated as pending working-tree churn.