[cubic/cuback] optionally use alpha=1 after loss

[kazuho]

Summary

• add quicly_context_t::cubic_loss_alpha_one
• when enabled, use alpha=1 for Cubic and Cuback packet-loss epochs below cwnd_prior
• retain the existing Reno-friendly alpha for ECN epochs
• preserve the setting across CC initialization, reset, and switching
• add unit coverage for Cubic, Cuback, ECN behavior, and policy switching

Simulation results

Each point below contains 60 paired runs, with the same randomized flow starts for stock, alpha-one, and Pico. Each run used a 10-second warmup followed by a 30-second measurement interval. Pacing and Rapid Start were enabled, ECN was disabled, and bottleneck utilization was approximately 100%.

Values are the tested flow's mean bottleneck share, with per-run SD in parentheses.

Against one stock Cubic contender

Fair share is 50%.

Profile	Tested CC	Stock	Alpha-one	Pico
BDP20	Cubic	50.26% (2.25)	51.12% (2.16)	53.82% (1.93)
BDP20	Cuback	48.88% (2.30)	51.44% (1.98)	53.84% (2.22)
BDP40	Cubic	50.82% (3.55)	53.25% (2.97)	53.17% (2.83)
BDP40	Cuback	48.24% (3.90)	52.16% (3.06)	53.71% (2.34)

Paired alpha-one minus stock differences, with 95% confidence intervals:

Profile	Tested CC	Difference
BDP20	Cubic	+0.86 points [0.03, 1.68]
BDP20	Cuback	+2.56 points [1.77, 3.34]
BDP40	Cubic	+2.43 points [1.30, 3.57]
BDP40	Cuback	+3.92 points [2.75, 5.09]

Against three stock Cubic contenders

Fair share is 25%.

Profile	Tested CC	Stock	Alpha-one	Pico
BDP20	Cubic	24.83% (1.58)	27.22% (1.49)	28.44% (1.84)
BDP20	Cuback	24.11% (1.52)	26.72% (1.53)	28.32% (1.59)
BDP40	Cubic	25.03% (1.45)	27.69% (1.72)	28.53% (1.68)
BDP40	Cuback	23.85% (2.13)	27.57% (1.59)	28.74% (1.63)

Paired alpha-one minus stock differences, with 95% confidence intervals:

Profile	Tested CC	Difference
BDP20	Cubic	+2.39 points [1.92, 2.86]
BDP20	Cuback	+2.61 points [2.09, 3.13]
BDP40	Cubic	+2.66 points [2.12, 3.21]
BDP40	Cuback	+3.72 points [2.99, 4.45]

In the three-contender experiment, alpha-one closed 62% to 76% of the stock-to-Pico share gap. The BDP20 Cuback runs were almost entirely Reno-friendly: stock recorded four non-friendly/crossover updates in total, and alpha-one recorded one. The lower-end bandwidth distributions were unimodal, without steep secondary peaks.

The heavily congested profile was excluded from the roofline tables because it was predominantly Cubic-controlled and Pico was less aggressive than Cubic there.

Tests

• cmake --build build/default -j2
• build/default/test.t