100E100M high-rate classes fail cwnd-settle before mouse latency probe

[psaab]

Finding

After #1353/#1358/#1361 and the #1364 drain-timeout branch, the
canonical 100E100M matrix can validate the default 5201 / 1 Gbps class,
but the first higher-rate class tried, 5202 / 10 Gbps, never reaches the
mouse probe window.

Artifact

/tmp/xpf-100e100m-exact-5202-10g-post1364-20260516-083030

Command shape

ELEPHANT_PORT=5202 \
SHAPER_BPS=10000000000 \
MOUSE_LATENCY_CELLS=$'0 100\n100 100' \
MOUSE_LATENCY_GATE_ELEPHANTS=100 \
MOUSE_LATENCY_GATE_MICE=100 \
MOUSE_LATENCY_GATE_PERCENTILE=p999_us \
MOUSE_PROBE_CONNECTION_MODE=persistent \
MOUSE_PROBE_MIN_INTERVAL_MS=20 \
./test/incus/test-mouse-latency-matrix.sh /tmp/xpf-100e100m-exact-5202-10g-post1364-20260516-083030

Result

N=0,M=100:   10/10 valid, representative p99.9 7068 us
N=100,M=100: 0/15 valid, all INVALID-cwnd-not-settled
verdict: INSUFFICIENT-DATA

This is not a mouse-latency PASS/FAIL. The loaded high-rate cell
invalidates before the probe driver starts because the elephant workload
does not satisfy the current 20-second cwnd-settle gate.

Interpretation

The 1 Gbps class evidence is valid and useful, but it does not prove the
high-rate classes. At 10 Gbps with 100 elephant streams, either:

• the source TCP workload genuinely needs a longer/adaptive settle period,
• the dataplane/scheduler is unstable enough that the settle gate is
  correctly rejecting it,
• the gate is too strict for high-rate 100-flow cells and needs richer
  evidence than the current pass/fail marker.

Acceptance

• Add explicit high-rate settle diagnostics to the artifact: final
  settle-window aggregate, per-flow min/median/max, retransmits, and
  reason threshold.
• Decide whether high-rate 100E100M should use an adaptive/longer settle
  budget or should fail as unstable.
• Produce a gate-grade 5202 / 10 Gbps exact artifact that reaches the
  mouse probe, or document why that shape is not a valid mouse-latency
  qualification target.
• Only after exact reaches the probe, repeat with
  MOUSE_COS_SURPLUS_SHARING=1.

Related

• 100E100M: surplus-sharing fails p99.9 mouse latency gate while exact passes #1359 tracks the product issue where default 5201 / 1 Gbps
  surplus-sharing creates mouse progress collapse / high p99.9 while
  strict exact passes.
• Bound mouse latency probe drain by deadline #1364 bounds probe writer.drain() and was used for this run.