Add BBRv3 as a selectable congestion control option

[endel]

Summary

Implements a faithful BBRv3 (draft-cardwell-iccrg-bbr-congestion-control-03) alongside the existing NewReno and Cubic, selectable via ConnectionConfig.congestion_control = .bbr. Default remains Cubic — BBR is opt-in, with zero per-packet overhead when not selected.

Architecture follows picoquic's bbr.c structure: three-stage onAckBatch pipeline, seven-state machine with four event-driven ProbeBW sub-states, full lower-bound dynamics, BBRLossThresh-gated inflight_hi, ECN EWMA, ACK aggregation, recovery state.

What's in here

Core:

• src/quic/bbr.zig (~1080 lines): Bbr struct, state machine, model estimators, pacing/cwnd computation.
• src/quic/delivery_rate.zig (~210 lines): RateSampler per RFC 9002 §B / draft-cheng-iccrg-delivery-rate-estimation.
• src/quic/congestion.zig: Algorithm enum + CongestionControl tagged union over NewReno/Cubic/Bbr with inline else dispatch; AckContext extended with prior_bytes_in_flight, newly_acked_bytes, newly_lost_bytes, largest_acked_pn; Pacer.setPacingRate(bps).
• src/quic/ack_handler.zig: per-packet snapshot fields on SentPacket; PacketHandler.rate_sampler + rate_sampling_enabled flag.
• src/quic/connection.zig: union dispatch; cc_algorithm field preserved across migration; rate-sampling enabled only when .bbr.

Docs:

• SPEC/9002_BBR.md: architecture, state machine, caveats.
• SPEC/STATUS.md: row added to RFC 9002 table.

State machine

Startup ──bw plateau / loss > BBRLossThresh / RTT excess─▶ Drain
Drain   ──inflight ≤ BDP─▶ ProbeBW_Down

ProbeBW_Down   ──inflight ≤ (1-headroom)·BDP─▶ ProbeBW_Cruise
ProbeBW_Cruise ──≥1 round elapsed─▶ ProbeBW_Refill
ProbeBW_Refill ──one round (resets bw_lo, inflight_lo)─▶ ProbeBW_Up
ProbeBW_Up     ──inflight ≥ inflight_hi / >2 rounds─▶ ProbeBW_Down

(any non-Startup state) ──min_rtt stale (5s)─▶ ProbeRTT
ProbeRTT       ──200ms drained─▶ ProbeBW_Down (or Startup if !filled_pipe)

Performance

	rps median (loopback HTTP/3)	Δ vs main
Pre-BBR main (bb61b7e)	21,010	baseline
This branch (default Cubic)	20,995	−0.07%

Within run-to-run loopback variance (5–10%). bench-codec identical within sub-ns noise. Per-packet rate-sampler overhead is gated behind rate_sampling_enabled so non-BBR connections pay nothing.

Caveats (documented in SPEC/9002_BBR.md)

• BBRRaiseInflightHiSlope uses a per-round approximation rather than the draft's per-ack slope.
• CRUISE uses simple "≥1 round elapsed" rather than picoquic's randomized timer + Reno coexistence quota. Acceptable on links not sharing bottleneck with Reno; revisit if fairness matters.
• No qlog instrumentation yet for BBR-specific state.
• Interop binaries don't yet have --cc bbr CLI flag (default Cubic).

Test plan

• zig build test — 551/551 pass (including 19 new BBR tests + 6 union-dispatch parity tests + 4 sampler tests + 3 connection-level algorithm-selection tests).
• zig build — all 20+ executables build clean.
• bench-codec (pure CPU) — identical to main within sub-ns noise.
• bench (loopback HTTP/3) — median rps within 0.07% of main.
• Cross-impl interop with .bbr selected (deferred — needs --cc bbr flag on interop binaries first).
• qlog capture from a long transfer to verify state transitions Startup → Drain → ProbeBW (deferred — needs qlog instrumentation).