start.sh: detect the fabric port on each Spark

[data-angel]

Same fix as Enntity/sparkglm#45 (reported in Enntity/sparkglm#44): this launcher had the same hardcoded port.

start.sh hardcoded the fabric port to rocep1s0f0, so a pair cabled on the ConnectX-7's second port (enp1s0f1np1) failed with no hint. .env.example already claimed the port was detected.

Now, on each Spark:

• start.sh lists the RoCE ports whose link is up and that have an IPv4 address.
• It picks the pair of ports on the same subnet, which is the cable between the two Sparks.
• It prints the choice and how to override it:
  == fabric: rocep1s0f1 (enp1s0f1np1, 192.168.100.1) here, rocep1s0f1 (enp1s0f1np1, 192.168.100.2) on <worker>; set FABRIC_HCA in .env to choose another port
• If no port pair shares a subnet, it falls back to the first active port and says to set FABRIC_HCA.
• FABRIC_HCA in .env still overrides both ranks. The engine still adds the sibling device behind the other PCIe domain of whichever port is chosen.

Tested:

• The selection logic on: a cable on port 1, mixed ports, two cables on one Spark, and no shared subnet.
• Real detection on our pair, where one Spark has two active ports (port 0 to the other Spark, port 1 to a third host). It picked port 0.

install/ is unchanged, so the image tag stays 656234c336af. The manual-install doc now mentions --fabric-hca rocep1s0f1 for the second port.

🤖 Generated with Claude Code