saage remote: list + cleanup commands, ps progress, terminate unregisters target

[cgpadwick]

Problem

Targets in credentials.toml accumulate forever: add-target appends, nothing removes them — even terminate leaves the registration behind. The result is a long list of dead names in every error message, and saage remote ps silently probing each dead box for a full ConnectTimeout, which reads as a hang.

Changes

• saage remote list — table of registered targets (name, dest, $/hr, per-target key). Pure local, no network.
• saage remote cleanup — interactive pruning: one remove <name> (user@host)? [y/N] prompt per target (default N, sorted order). Warns when an unfinished run in the local ledger references the target. Optional --check ssh-probes each target first and shows reachable/unreachable as information only — offline every box looks dead, so reachability never decides anything. Prints a reminder after removals that removing a target does not terminate the box.
• creds.remove_target() — deletes the [targets.<name>] section by text splice so the rest of the file stays byte-identical (comments and [storage]/[lambda] sections survive; a TOML re-emit would strip comments). Never deletes key files (per-instance Thunder keys are unrecoverable). Preserves 0600.
• saage remote terminate — now also unregisters the matching target, so spawned boxes stop becoming permanent clutter.
• saage remote ps — narrates each probe (checking <name> (<dest>)… ok (N sessions) / unreachable) so a wall of stale targets reads as progress instead of a hang.

Design notes: docs/superpowers/specs/2026-08-01-remote-target-cleanup-design.md.

Testing

• 19 new offline unit tests (tests/remote/test_target_cleanup.py, additions to test_creds.py): splice removal edge cases (middle/last section, comment preservation, 0600, key files untouched), cleanup prompt loop with scripted input, active-run warning, --check output, ps progress lines with a stubbed SshTarget, terminate unregistration.
• tests/remote/: 103 passed, 5 skipped (ssh/live markers).
• Full suite: 415 passed; the 5 failures (test_skills, test_agent, test_checkpoint warning-capture tests) also fail on clean master and are unrelated.
• Manual smoke: list → scripted cleanup → list confirms removal and billing reminder.

🤖 Generated with Claude Code