Add who/free, launcher attribution, and --mine/--orphaned scopes

[danfry1]

Summary

Coding agents now start most local dev servers, often in the background and across several concurrent sessions, which makes a running server hard to identify. This change makes marina answer what is this, and who started it, and adds port-first commands for the EADDRINUSE moment.

marina who / marina free

• who <port>… names whatever holds a port: a dev target (project · tool · pid · uptime · branch · cwd · launcher) or, for anything outside curation, the process name + pid. Exit 1 when free.
• free <port>… stops the dev target on a port and waits until it is actually released. Idempotent (marina free 3000 && pnpm dev re-runs safely); refuses to touch non-dev holders such as macOS ControlCenter on :5000. Ports only — no substring matching.

Launcher attribution

• Every target records its launcher: an agent session (claude, codex, opencode, cursor-agent, …), an editor, a terminal, or detached.
• Resolved from the live parent chain; when a server has been reparented after &/nohup, falls back to an allowlist of agent env markers (CLAUDE_PID, CLAUDE_CODE_SESSION_ID, CLAUDECODE, AI_AGENT). A server whose agent session has ended is flagged orphaned (claude·ended).
• TUI: VIA column (dropped after BRANCH/UP on narrow panes), launcher detail in the status line and inspect panel (with claude --resume <id>), / filter matches launcher (/claude, /ended).
• JSON: additive launcher object {kind, name, pid, alive, orphaned, session, cwd}.
• --mine / --orphaned scopes on ls/kill/restart/url: an agent can run marina kill --mine to stop exactly the servers it started, and marina kill --orphaned cleans up after finished sessions.

Resolution and verbs

• Labels see through corepack's pnpm-native shim and node-hosted bin scripts (pnpm exec vp dev → vp dev); python -m http.server → http.server.
• A server whose cwd is $HOME is named after its tool rather than the username.
• CLI kill polls for exit instead of always sleeping the grace period (free/kill ~1.5s → ~30ms typical).

Privacy

Only the four allowlisted env keys are extracted; no other variable is kept, displayed, or serialized, and session ids are validated as hex/dashes. Claude Code's ~/.claude/sessions/* files are intentionally not read because they contain auth tokens. README privacy section updated accordingly.

Testing

• cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo test --all-targets: 101 tests pass (17 new), covering the env allowlist, launcher classification without over-matching (codesign ≠ code), chain vs. env precedence, orphan detection, detached vs. unknown supervisors, --mine session matching, port parsing, who/free exit codes against real sockets, label resolution, and TUI rendering of the VIA column / inspect panel / filter.
• Verified manually on macOS against live servers: agent-launched Vite servers attribute to their exact Claude Code session; a nohup-backgrounded server still attributes via env; a server with a dead CLAUDE_PID shows claude·ended and is stopped by kill --orphaned; --mine excludes servers from other sessions; the TUI renders correctly at 130 columns.
• Not yet verified on Linux beyond CI.

Session view and orphan notices (second commit)

• Session view (v): groups rows by the agent session / editor / terminal that started them, across projects — claude·fal-app, claude·ended, tmux, detached. K on a session header stops everything that session started.
• Orphan notices: when an agent session ends and leaves servers running, a single notice per session — ⚠ claude session in ~/dev/app ended — left 2 servers running (480MB) — in the status line (20s) and as a desktop notification. Disable the latter with [notify] desktop = false. Orphans present at startup get a status-line count only.
• Fix: agent env markers without CLAUDE_PID are no longer treated as an ended session (no false orphan alarms).

Verified live on macOS: a stand-in agent process with two env-attributed servers grouped under claude·agent-pr… (2); killing the agent moved them to claude·ended (2) and raised the notice; /ended + K on the session header stopped exactly those two servers while other sessions' servers were untouched. 109 tests (8 new: session grouping and disambiguation, session-view kill scope, once-per-session notices, startup orphans, pid-less markers, header truncation, AppleScript escaping, notify config).

Multi-agent detection (third commit)

• Marker parsing is table-driven and covers Codex (CODEX_THREAD_ID, CODEX_SANDBOX), opencode (OPENCODE_SESSION_ID, OPENCODE), Copilot, Cursor, Gemini, Amp, Goose, Cline, Roo, Qwen, Pi, Augment, Crush and Kiro, with AI_AGENT/AGENT as a fallback. Sources: Codex exec_env.rs, opencode's bash tool (read from the shipped binary), cross-checked against is-ai-agent.
• --mine and session grouping work for Codex and opencode via their thread/session ids; inspect shows codex resume <id> / opencode --session <id>.
• Editor-hosted agents (Cursor agent, Copilot agent mode) are attributed to the agent rather than the editor; terminals are deliberately not upgraded.
• Only Claude Code exports a pid, so only Claude sessions can be proven ended; other agents' servers are never flagged orphaned.
• Fix: the anchor climb stops at any launcher. Found by running a server through the real codex sandbox command: with no shell in between, the Codex process itself became the target's anchor (row labelled codex, Codex in the kill set).

Verified live: a listener started via codex sandbox reads nc · via codex (pid …); backgrounded servers carrying Codex / opencode markers read via codex (session 0199a213) / via opencode (session ses_3b0f); ls --mine run with each agent's session marker returns only that session's server. 117 tests.