Add marina port: stable, collision-free ports per project/worktree

[danfry1]

Summary

Parallel coding agents working in parallel git worktrees all default to the same dev port (usually :3000), so the second server fails with EADDRINUSE or silently lands on a random port. marina port gives every project root its own stable port:

pnpm dev --port $(marina port)        # each worktree: its own port, every time
PORT=$(marina port api) cargo run     # several services in one tree
marina port --list                    # all assignments (listening / idle / gone)
marina port --release [name]          # give one back

Design

• Key: project root of the canonical cwd (each worktree, each monorepo package) + optional service name.
• Stable: first choice is FNV-1a(key) within the range (default 3100–3999, [ports] range in config); the lease is recorded in $XDG_STATE_HOME/marina/ports.json, so re-runs return the same port.
• Collision-free: the read-modify-write runs under an exclusive flock and saves via temp-file rename; probing skips ports leased to other keys and ports busy with a foreign listener (netstat + bind probe). A busy port whose listener is a dev target inside this root is the project's own server and is kept; a lease taken over by another process moves (stderr note, "status": "moved" in --json).
• Self-cleaning: leases for roots that no longer exist (removed worktrees) are dropped on every call.
• ports::allocate is pure (I/O and busy probe injected); stdout is only the port number so $(marina port) composes.

Docs

README (usage + privacy: ports.json is now a second user-triggered write), DESIGN, CONTEXT glossary, and the agent skill — agents are told to use --port $(marina port) rather than hard-coding ports and to --release when done with a worktree.

Testing

• cargo fmt --check, clippy -D warnings, cargo test: 127 tests (10 new) — stability across calls and fresh registries, uniqueness across 200 roots × 2 names, foreign takeover → move, deleted-root GC, range exhaustion, release, name validation, range config, CLI usage errors, and 8 threads allocating concurrently through the real locked store (no duplicates, no lost updates).
• Live on macOS with real git worktrees (isolated XDG_STATE_HOME): repeat calls stable; main / feat-a / feat-b / api all distinct; 10 concurrent marina port processes → 10 unique ports; the project's own server keeps its port; a foreign server taking feat-b's port moves it (3978 → 3979); removing the feat-b worktree drops its lease; --release frees and then reports nothing to release (exit 1).