feat(compat): SRT receive-profile validation — A/B matrix, FEC connect-matrix, srt-sink TS metrics

[andrescera]

What

Adds the srtla-side validation harness for the SRT receive-profile capability:

• srt-sink TS-continuity metrics (task 5): ts_packets, ts_sync_errors, ts_cc_errors, pkt_rcv_loss, pkt_rcv_drop, pkt_retrans — TS-continuity signal replaces bytes-only as the regression gate. --retransmitalgo and --reorderfreeze flags added.
• FEC connect-matrix (task 7): empirically proves the one-sided SRT FEC negotiation — a fec-accept listener accepts non-FEC callers plain (no reject), so FEC rides L1 with no separate listener needed. 4 cases, all pass. Documented in docs/COMPATIBILITY.md.
• A/B profile validation matrix (task 6): paired baseline-vs-profile comparison for all 4 non-FEC profiles (Balanced/Low-Latency/Resilient/Classic) under reorder stress. Equal-gate: disconnects==0, ts_sync_errors==0, ts_cc_errors≤baseline, goodput≥99%, p95-drop≤baseline, wire-amp≤1.10×. All 4 PASS. LOSSMAXTTL sweep {30,200,1000} all PASS; recommended cap=30.
• AGENTS.md updated with new compat harness capabilities.

Why

The A/B matrix is the regression gate that proves the SRTO_REORDERFREEZE option (CERALIVE/srt) does not degrade stream quality vs the patched BELABOX baseline. The FEC matrix proves the architectural decision that FEC rides L1 (no separate listener). Both are required before the receiver cutover.

How to verify

cmake -B build -DBUILD_COMPAT_TESTS=ON && cmake --build build -j
# FEC matrix
bash tests/compat/scenarios/fec-connect-matrix.sh
# A/B matrix (requires pre-built libsrt installs or BASELINE_LIBSRT/FREEZE_LIBSRT envs)
bash tests/compat/scenarios/profile-validation-matrix.sh

Risks

Low. All changes are under tests/compat/ (BUILD_COMPAT_TESTS=ON only) — no production code path touched. Default cmake build is unaffected (Rule E verified: existing compat pairs still pass).