feat: add RFC 3264 call hold / resume

[wavekat-eason]

Adds a first-class signaled call-hold primitive so consumers can put a confirmed call on hold (and resume it) via an RFC 3264 re-INVITE, instead of hand-rolling SDP re-INVITEs or faking hold with a local audio gate the far end never learns about. Per the crate's placement rules, SIP signaling belongs here.

What's added

• sdp::MediaDirection — SendRecv / SendOnly / RecvOnly / Inactive, with attr() (wire token), parse(), and responding() (the direction to answer a peer's offer with). Defaults to SendRecv.
• sdp::build_sdp_with_direction() — the existing build_sdp body with an explicit direction line and an o= session version; build_sdp delegates to it (SendRecv, version 0), so its output is unchanged.
• sdp::RemoteMedia::direction — the parsed direction of a remote SDP (default SendRecv when absent), so a consumer can also detect a peer-initiated hold.
• hold::reoffer_media() — sends the directional re-INVITE over any SessionDialogOps dialog (reusing the session-refresh seam) and returns the re-parsed answer.
• hold::HoldHandle — a cheap, cloneable handle (both inner dialog types are Clone) so a consumer can snapshot it, release whatever lock guards its call table, and only then await the re-INVITE round-trip. It carries a call-scoped Arc<AtomicU64> o= version counter so successive hold/resume re-offers advance one monotonic series.
• AcceptedCall::set_hold / hold_handle, AcceptedDial::set_hold / hold_handle — convenience wrappers. They own the version counter (seeded to 0) and clone it into every handle.

The re-offer keeps the same connection address, port, and codecs — only the a= direction moves — so the dialog and any session timer survive the transition.

RFC 3264 §8 o= version (hold→resume interop)

The o= origin line keeps a fixed session id but its version must increment on every re-offer that changes the session. An early revision hardcoded o=wavekat 0 0, so hold and resume both carried version 0 — and a carrier SBC (2talk) accepted the hold but rejected the resume with 500 Server Internal Error, then BYE'd the call. The version is now threaded through build_sdp_with_direction and sourced from the per-call counter: initial offer/answer 0, first re-offer 1, next 2, … Because hold_handle() mints a fresh handle per request, the counter lives on AcceptedCall/AcceptedDial and is cloned in — a per-handle counter would reset and reintroduce the bug.

Out of scope (consumer's job)

What plays on the RTP stream while held (silence, music-on-hold) is audio, not signaling. Answering a peer-initiated hold re-INVITE is left to the consumer too — this crate supplies RemoteMedia::direction + MediaDirection::responding() but does not auto-answer.

Tests

sdp.rs covers the direction round-trip (all four directions), the absent-attribute default, responding(), and that the o= version tracks the fed session_version. hold.rs unit-tests reoffer_media against a scripted SessionDialogOps mock: asserts the offer's direction/port and o= version, parses the answer, maps a bare 2xx to Ok(None), errors on a non-2xx / unconfirmed dialog, and checks the version counter yields a strictly increasing 1, 2, 3, … series. No live SIP server needed. cargo fmt/clippy/test/doc all clean.

Design note: docs/08-call-hold.md; coverage updated in docs/RFC-COVERAGE.md.

🤖 Generated with Claude Code

[wavekat-eason]

Superseded by #43 — hold/resume is reimplemented on the in-house engine there (Call::set_hold / is_held). Closing in favor of the consolidated 0.1.0 PR.