DecidePort() returns 0 when source and destination IP addresses are
equal (or when parsing fails). The previous check 'result < 0' mapped
this case to Interface1, making the assignment non-deterministic and
asymmetric. Changing to '<= 0' ensures equal addresses always go to
Interface0, providing a stable deterministic default.

std::runtime_error was constructed but immediately discarded without
being thrown. When the histogram has not been initialized the function
silently continued, leading to undefined behaviour on the empty vector
access below. Add the missing 'throw' keyword.

The loop condition 'windowIndex <= end_window_idx / WINDOW_SIZE + 1'
allowed windowIndex to reach nOfWindows when end_window_idx was the
last valid bin, which is one past the end of windowsAcc (size
nOfWindows). The spurious '+ 1' served no purpose since the inner body
already guards the write via 'if (pkts > 0)' and GetHistogramBin()
returns 0 for out-of-range indices. Remove the '+ 1' to keep the loop
strictly within bounds.

CreateInitialPopulation() called fut.wait() on the worker futures,
which blocks until completion but silently discards any stored
exception. If a worker throws, the population is left in an incomplete
state and UpdateFitnessStats() runs on bad data. Switching to
fut.get() re-throws the stored exception to the caller.

…Worker()

std::uniform_int_distribution<uint64_t>(0, profileSize) is inclusive on
both ends, so it can generate an index equal to profileSize. Since
genotype has size profileSize (valid indices 0..profileSize-1), this is
an out-of-bounds access causing undefined behaviour. Change the upper
bound to profileSize - 1.

The 'bidir' member had no default initializer, so default-constructing
SMSubnetSegment (e.g. from pybind11 py::init<>()) left it with an
indeterminate value. Contains() reads bidir unconditionally, resulting
in undefined behaviour. Initialize to 'false' (unidirectional) as the
safe default.

sysconfig.get_path("stdlib") returns the standard library directory
(e.g. /usr/lib/python3.9), not the site-packages directory. The path
was then extended with a hardcoded "/site-packages/flowtest" suffix
which accidentally worked on some layouts but is not guaranteed.
Use sysconfig.get_path("platlib") which directly returns the compiled
extension modules site-packages directory, and os.path.join() for
safe path construction.

ParseSpeedUnit() applied 'multiplier * 8' before returning, turning
e.g. '100 gbps' into 800e9 instead of the correct 100e9 bits/sec.
Since gbps/mbps/kbps/bps units are already in bits per second, the
extra factor of 8 was wrong and inconsistent with the default
_linkSpeedBps value of 100e9 (100 Gbps). Remove the '* 8' so the
function returns the correct bits-per-second value.

…ead of sum

The C++ implementation (CalcMinGlobalPacketGapPicos) computes the
minimum inter-packet gap as the wire transfer time PLUS the configured
min_packet_gap: 'minGap = transferTime + minPacketGapNanos'. The Python
test helper incorrectly returned min(transfer_nanos, min_packet_gap_ns),
which can be smaller than the actual minimum gap enforced by the code.
This made the test weaker than intended and could mask regressions.
Change to use the sum to match the implementation.

rte_eth_dev_set_mtu() and the devInfo min/max_mtu validation both
operate on the L3 payload MTU (e.g. 1500 for standard Ethernet), but
_MTUSize (default 1518) represents the full Ethernet frame size
including the 14-byte header and 4-byte FCS. Passing the frame size
directly caused an 18-byte discrepancy. Define ETH_OVERHEAD = 18 and
subtract it when calling rte_eth_dev_set_mtu() and when comparing
against devInfo min/max_mtu.