feat(gc): sandy --gc — unified Docker-resource reclaim (#36)

[rappdw]

Closes #36.

Problem

Sandy accumulates Docker resources during normal use with no reclaim path: dangling <none> images from same-tag rebuilds, orphaned per-project/skill images, and dead-owner containers left by SIGKILL/OOM where the cleanup() trap never fired. Networks were already handled (--prune-orphans); nothing else was.

What this adds

sandy --gc — a global, workspace-independent fast-path command (same pre-preflight family as --print-state/--prune-orphans; needs only a reachable Docker daemon) that reclaims all four resource types:

• dead-owner containers — sandy-* / sandy-proxy-* whose owning session is dead
• orphan networks — folds in the existing _sandy_reap_orphan_networks
• orphaned per-project / skill images — sandy-project-<name> / sandy-skills-* no longer referenced by a sandbox/config
• dangling <none> sandy images — scoped by a new sandy.managed=1 build label so a stranger's dangling image is never touched

--dry-run / --yes mirror --update-sessions (nothing-to-do → exit 0; non-TTY with something to reclaim and no --yes → exit 1; TTY → y/N). Reap order is containers → networks → images so docker rmi (no -f) actually frees. Two new full-mode-only --print-state keys — dangling_images, orphaned_containers (null in light mode, zero added light-mode spawns). schema_version stays 1.

Safety — never reap a live session's resource

• Container liveness reuses the "container is truth only with a live inner tmux session" rule (D6/DEC-B/--stop), never trusting sandy.daemon_pid — so the D9 reboot case (a container --restart-resurrected with a dead supervisor pid) stays alive via a tmux has-session probe.
• Agent-vs-proxy is discriminated by IMAGE, not name prefix, and a proxy's liveness follows its paired agent (two-pass classifier) — keeps a live daemon's proxy when its supervisor was killed, and stops a workspace literally named proxy from being mis-stripped.
• The reaper's daemon probe uses the same 5×/sleep 1 mid-startup retry as --start so a concurrent launch isn't reaped out from under itself.

Process

Plan (fable) → implement (Sonnet 5) → adversarial verify (Opus 4.8). The verify pass caught two live-session reap paths (proxy strand of a supervisor-less daemon; proxy-named workspace) + two majors (mid-startup race; skills in-use set ignoring host-global SANDY_SKILL_PACKS). All four fixed and re-verified: SHIP — no new false-positive paths, tests genuinely bind.

Tests

run-tests.sh §79 — 49 assertions: structural + a stateful stubbed-docker harness covering the D9-alive-survives, unlabeled-dangling-untouched, proxy-paired/proxy-named, orphan-proxy, mid-startup-retry, and host-config-skills cases. Real reclaim against live Docker is deferred to a maintainer-run acceptance harness (like other Docker-runtime features).

Notes

• --prune-orphans kept unchanged as a documented network-only subset (shares the lister/reaper, can't drift).
• Retroactive gap: images built by pre-1.3.0 sandy lack the sandy.managed=1 label and won't be caught by the dangling-image step — self-heals as images churn.

Additive-only (1.3.0); sandbox forward-compat holds.

🤖 Generated with Claude Code