Wire psrpc bus compression into the message bus constructor

[paulwe]

psrpc v0.7.6 (psrpc#130) added opt-in gzip at the bus boundary, and protocol (#1771) surfaced it as rpc.PSRPCConfig.Compression with a BusOptions() conversion. livekit-server (#4844), egress and ingress have all landed the wiring.

sip was the straggler, which mattered in two ways: it had no psrpc: config block at all, so the setting was unreachable here; and sip on psrpc v0.7.5 cannot decode a compressed payload — psrpc's subscription read loop drops undecodable messages without an error. Bumping sip is therefore a precondition for enabling compression anywhere on the shared bus.

psrpc:
  compression:
    quality: 6
    threshold: 1024

Compression is off by default

quality defaults to 0. Enabling it is a two-stage operator action: roll psrpc v0.7.6+ out to LiveKit server, egress, ingress, SIP and any agent workers first, then raise quality at the publishers. The README says so at the setting, along with a note that the other PSRPCConfig keys (max_attempts, timeout, backoff, buffer_size) parse but are unused here — NewConfig uses non-strict yaml.Unmarshal, so they'd otherwise silently look effective.

max_decompressed_size defaults to 0, unlimited, so nothing starts being dropped by a cap nobody set.

Shape

Follows egress (cmd/server/main.go:117) and ingress (cmd/server/main.go:132) exactly: a PSRPC rpc.PSRPCConfig field seeded with rpc.DefaultPSRPCConfig in the NewConfig literal before yaml.Unmarshal, so a partial override keeps the seeded threshold.

Seeded in NewConfig only, not Init(). A zero-valued rpc.PSRPCConfig is behaviourally identical to the old no-options bus — newCompressor returns nil for Quality <= 0, getBusOpts re-seeds Threshold to DefaultCompressionThreshold when it's <= 0, and MaxDecompressedSize: 0 is unlimited on read — so the two test paths that bypass NewConfig need no rescue. Init() also runs after NewConfig, so it couldn't assign wholesale without clobbering YAML values.

Both test bus call sites are wired too, so grep NewRedisMessageBus\|NewLocalMessageBus finds no stragglers. Neither changes behaviour.

Dependency bump

protocol → v1.51.1-0.20260905133529-a4f4b5c0c23f and psrpc → v0.7.6, the same revisions egress and ingress are on.

The protocol range carries one behavioural rider sip actually uses: TEL-955: fix potential panic in sip, which guards an empty INVITE user part in EvaluateDispatchRule. A fix sip wants. The rest are inert here — the grpc v1.83.1 security bump is already satisfied, sdp: harden SDPFragment.Unmarshal touches a package sip imports in zero files, and the billing-reporter and agent-simulation commits are additive protobuf in packages sip doesn't import.

go mod tidy also drops a stale golang.org/x/mod indirect requirement. That drift predates this change — running tidy on the unbumped deps produces the identical removal — but it rides along rather than being hand-restored.

Testing

• go build ./..., go vet ./... and go test -race ./pkg/... pass; both test/ packages compile.
• Verified the YAML reaches the bus: a partial override yields quality=6 threshold=1024 max=4096. threshold surviving at 1024 rather than resetting to 0 is what proves the seed-before-unmarshal ordering. Defaults yield quality=0 threshold=1024 max=0.
• No pkg/config/config_test.go — this package has never had tests and this change doesn't introduce the first one. Worth knowing that leaves the seed-before-unmarshal ordering unguarded: moving the PSRPC: line below yaml.Unmarshal would silently reset threshold to 0 and start compressing every small message. Happy to add two short tests if reviewers want it closed.

Note on formatting

pkg/config/config.go and cmd/livekit-sip/main.go are already non-gofmt-clean on main, in regions untouched here (the MediaTimeout/RTPDraining alignment block, and the long one-liner at main.go:102). Running the formatter would have swept that unrelated churn into this diff, so I didn't — every line added here is independently goimports-clean. CI runs only staticcheck, no formatter, which is why the drift survives. Happy to fix it separately.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.