test(compat): add ceralive-srtla-send-rs blocking pair; docs: ADR-003 Rust sender adoption

[andrescera]

Affected repo & language: srtla — Bash (compat harness/scenarios), YAML/Python (matrix registry), Markdown (docs/ADR)

What

Registers the CERALIVE Rust fork (srtla-send-rs, the ADR-003 cutover sender) as a blocking compatibility pair against our C srtla_rec, and commits ADR-003. The compat harness gains a forkbin sender kind that runs the fork's pre-built release binary (resolved from SRTLA_SEND_RS_BIN, or a srtla_send_rs on PATH). The five behavioral scenarios are made sender-agnostic so the same assertions validate both the C sender and the Rust fork.

Why

ADR-003 adopts the Rust fork as the shipping srtla_send via direct cutover. The cutover gate requires the fork to pass the blocking compat matrix against the C receiver, and the failure/stress scenarios must keep working once the C sender is retired from the .deb. The fork uses different log wording than the C sender and is silent without RUST_LOG, so the scenarios had to learn both vocabularies without weakening any pass criterion.

How to verify

• cmake -B build -DBUILD_COMPAT_TESTS=ON && cmake --build build -j
• Fetch the fork binary: gh release download v1.0.0 -R CERALIVE/srtla-send-rs --pattern "*amd64*.deb", verify SHA256 803f0ed4…, extract /usr/bin/srtla_send.
• SRTLA_SEND_RS_BIN=<that binary> bash tests/compat/run-matrix.sh --tier blocking → all 7 pairs PASS (incl. ceralive-send-rsxours).
• Falsifiable: SRTLA_SEND_RS_BIN=… bash tests/compat/run-matrix.sh --pair ceralive-send-rsxours --scenario port-mismatch → FAIL (exit 1).
• Scenarios vs the fork (point --build-dir at a dir whose srtla_send is the fork): receiver-restart, sighup-reload, link-drop, reorder-stress PASS.
• python3 tests/compat/validate-matrix.py → PASS (7 blocking pairs). ctest --test-dir build → 178/178 green.

Risks

• The Rust fork emits no logs unless RUST_LOG is set; each scenario's sender launch prefixes RUST_LOG="${RUST_LOG:-info}" (a no-op for the C sender). Removing that prefix would blind the scenario greps to the fork.
• jitter-stress's privileged netns/netem path is environment-sensitive and does not validate cleanly on every host for either sender (the C baseline also fails it on the QA host: no second link / zero throughput). Its 0-reaps criterion was left unrelaxed; the fork additionally exhibits socket-reconnection churn under ±200 ms jitter while keeping the bonded stream healthy (both links registered, disconnects==0). Tracked as a follow-up, not a cutover blocker.
• The matrix entry pins the fork by release tag/commit + .deb SHA256 (package Version is 3.0.0 from the upstream Cargo.toml); pin by SHA/commit, not version.

---

Checklist

• Docs updated (Rule A): AGENTS.md COMPAT section, docs/COMPATIBILITY.md ecosystem table, ADR-003 committed
• Started from updated main; branched off origin/main (Rule B)
• git grep -n '.omo' -- ':!.gitignore' returns nothing (Rule D)
• Tests pass; QA evidence captured (blocking matrix 7/7, falsifiability FAIL, ctest 178/178)