feat(net): support the '*' wildcard in --net-allow-bind

[congwang-mk]

Closes #133.

--net-allow-bind '*' allows the sandbox to bind any TCP port, including an ephemeral bind(0), while the rest of the network policy stays active. Previously the only way to approximate this was --net-allow-bind 0-65535, which installs 65536 individual Landlock rules.

• Sandbox.net_allow_bind becomes a BindPorts enum (Ports(Vec<u16>) or All); pre-1.0 hard break.
• Mixing * with port lists is rejected at build time, as is * in --net-deny-bind; repeating the bare wildcard is idempotent. Mutual exclusion with --net-deny-bind is unchanged.
• Enforcement: All drops LANDLOCK_ACCESS_NET_BIND_TCP from the handled access set (the same mechanism --net-deny-bind already uses), so no per-port rules are installed. Connect-side handling is untouched.
• C ABI cleanup (pre-1.0 hard break): the per-port entry points sandlock_sandbox_builder_net_allow_bind_port / net_deny_bind_port are removed in favor of spec-string entries sandlock_sandbox_builder_net_allow_bind(spec) / net_deny_bind(spec), matching how net_allow/net_deny already work. Port-spec grammar and the wildcard rule are parsed and validated in one place (the core builder); the Python and Go SDKs forward specs verbatim, and their local port/wildcard parsing (bind_all()/bind_ports()/deny_bind_ports(), Go ParsePorts/HasPortWildcard) is deleted.
• Wired through every surface: CLI, TOML profiles (allow_bind = ["*"]), the C ABI, the Python SDK (net_allow_bind=["*"]), and the Go SDK (NetAllowBind: []string{"*"}).
• Tests: parser/exclusivity/Landlock-mask unit tests; kernel-level integration tests (default denies bind with EACCES; wildcard allows fixed-port and port-0 binds; wildcard bind alongside connect enforcement under both the Landlock deny-all and the on-behalf bounded-net_allow enforcers); Python config/profile/end-to-end tests; a Go end-to-end wildcard test.

🤖 Generated with Claude Code

[dzerik]

Traced the Landlock path end-to-end — the bind-access semantics are exactly right: All drops only BIND_TCP from the handled mask, CONNECT_TCP and the net_wildcard derivation are untouched, no per-port rules are installed, and the on-behalf bind_deny_ports path stays uninvolved, so --net-allow-bind '*' alone still default-denies outbound TCP. The two mask-contract tests pin this. FFI/Python/Go wiring is faithful and the new spec-string C entry mirrors net_allow/net_deny for error-path ownership. LGTM overall.

A few non-blocking notes:

1. Doc drift — the old parse_bind_ports docstring ("expand into a sorted, deduplicated port list … Mirrors the Python SDK's parse_ports") is now stacked above parse_allow_bind_ports (sandbox.rs:2350), which returns BindPorts, while parse_bind_ports at :2373 is left undocumented. Worth moving the 3 lines down.
2. Wildcard predicate is implemented three times — parse_allow_bind_ports, Python bind_all(), Go HasPortWildcard. They agree today, but there's no shared test vector guarding drift; a divergence would let one binding forward a spec another rejects. A single cross-language vector would lock it.
3. The headline "rest of the network policy stays active" isn't exercised at runtime — test_net_allow_bind_wildcard_permits_any_bind only bind()s on base_policy() (no net_allow), so connect-enforcement-under-* rests entirely on the mask unit test. One case combining .net_allow_bind("*") with a bounded net_allow, asserting a disallowed connect() still fails EACCES beside a succeeding bind, would guard the actual claim.
4. Parity nit — Go has no end-to-end test for the * path (only HasPortWildcard); Python got a real bind test plus both exclusivity errors.

Nit: --net-allow-bind '*' '*' errors with "cannot be combined with port lists" — a repeated bare wildcard could just be idempotent.

[congwang-mk]

Thanks for your review. I have addressed them and refactored the FFI API together too.

[dzerik]

Re-read the update — all cleanly addressed, and asserting EACCES rather than tolerating an ECONNREFUSED in the connect test is exactly the right instinct. LGTM, thanks for the fast turnaround.