fix: unblock a real OpenPlay cast — discovery, encoder config, capture caps

[Developer1010x]

Found by running both binaries and actually casting. None of this is reachable from the test suite: the sender path needs a desktop portal and a human at the consent prompt, which is exactly why webrtc_loopback.rs uses a videotestsrc stand-in.

Three defects, each hidden behind the one before it.

1. The sender discarded every mDNS update

add_receiver was guarded by if !contains_key, so the first advertisement for a name won permanently. mDNS resolves addresses incrementally — the first event routinely carries only link-local addresses, and the routable ones arrive in later events for the same name.

The sender therefore held a set of fe80:: addresses, none dialable, because nothing advertises a zone index and one cannot be reconstructed (address.rs says so in its own doc comment):

Connecting to receiver url=wss://[fe80::3cec:9aff:fea6:cac0]:7290/
  ... IO error: Invalid argument (os error 22)
  ... and the same for all six link-local addresses
Could not reach the receiver on any advertised address

Worth noting this is the other half of #48. That change made the sender try every address instead of only the first; this one makes sure the reachable addresses are in the list to try.

Each event carries the full set ServiceInfo::get_addresses knows rather than a delta, so the newest is the most complete and replacing outright is right. The key is the service fullname and is stable across updates, so a row keeps its list position and a selection on it survives.

2. A hardcoded rate-control="cbr" brought the cast down

The va plugin builds that enum per driver — the type is literally GstVaEncoderRateControl_H264_LP_renderD128 — and populates it from what the hardware reports. A Tiger Lake iGPU offers only cqp on vah264lpenc:

rate-control : Enum "GstVaEncoderRateControl_H264_LP_renderD128" Default: 16, "cqp"
                  (16): cqp - Constant Quantizer

set_property_from_str panics on a nick the enum does not define, so the cast died immediately after the user pressed Allow:

property 'rate-control' of type 'GstVaH264LPEnc' can't be set from string 'cbr'

probe_best_encoder could not have caught this. It instantiates each candidate and instantiation succeeds — configuring the element fails one step later. set_enum_property now picks the first nick the driver actually defines, and keeps the element's default if none match, so an unusual driver costs some quality instead of the whole cast.

3. Nothing converted the portal's frames for the encoder

The capture source fed the hardware encoder directly, with a fixed framerate=30/1 imposed on a portal stream that advertises a maximum rather than a fixed rate. The intersection was empty:

pipewiresrc0: stream error: no more input formats
streaming stopped, reason not-negotiated (-4)

videoconvert widens the accepted formats to everything raw, and videorate is what actually produces the fixed rate the capsfilter asks for — without it that filter is a demand the source cannot meet.

Verified

Cast between the two binaries on one machine, driven by hand:

sender:   Sender pipeline started / Negotiation needed — creating offer
          Local description ready kind=Offer bytes=795
          WebRTC connection state state=Connected
          Media session connected — streaming
receiver: SDP offer accepted — answering
          Local description ready kind=Answer bytes=621
          WebRTC connection state state=Connected
          Using hardware decoder decoder="vah264dec"
          Video decoding chain linked (RGBA appsink)

This is the first OpenPlay cast confirmed working end to end — the project has carried a note that it never had been.

cargo test --all 323 passing / 0 failures; fmt and clippy -D warnings clean.

Known remaining defect

The displayed image is uniformly green. Connection, negotiation and decode are all correct, so this is a colour or buffer-format problem in the capture-to-encoder hop — most likely the portal handing out DMABuf that a CPU videoconvert cannot import, given pipewiresrc is built with always-copy=false. Being tracked separately; this PR is what gets a cast to run at all.

Also still unverified between two machines.

🤖 Generated with Claude Code