ci(cost-sweep): green on main with no arm ever placed and routed (XDC path + OPAD auto-placement) #896

What

tnf-cost-sweep.yml has been green on main while no arm was ever placed and routed since the per-arm XDC landed (W928, 8b165f5, 2026-08-19). Green there meant "nothing crashed", not "something was measured".

Evidence

Two defects, one behind the other:

1. Path. The XDC is written to the runner's /tmp/${TOP}.xdc; the container mounts only $PWD:/work, so nextpnr cannot open it.
2. Pins. With only IOSTANDARD set (the branch above writes the XDC under build/, so it gets past defect 1), nextpnr-xilinx auto-places an output on a GTP pad (OPAD_X0Y11) and placement fails.

Both artefacts carry status=routing-pending, which reads like a fabric limit. Neither run reached the router.

Fix

Branch fix/tnf-cost-sweep-xdc-path:

• the XDC lives under build/tnfcost/ (seen as /work/build/tnfcost/);
• new tools/sweep_pin_xdc.py gives every port bit a PACKAGE_PIN from the image's own package_pins.csv: general-purpose IOB sites only, clk on an MRCC P pin, bank 34 (the AX7203 DDR3 bank) excluded. Selftest 13/13. On a locally synthesized e3m33 it pins 123 pads with no pin used twice;
• an arm whose log never shows Info: Routing.. or Info: Running router2 is recorded as harness-failure and fails the job. Real routing failures stay routing-pending and do not. Negative control: both saved OPAD logs from run 34448747524 contain neither marker (0/0), so they would now fail the job.

Verification run on the branch: to be linked in the PR.

Not measured

• Whether the pinned arms route, and at what Fmax -- that is the verification run.
• Board behaviour. These pins are not the AX7203's; an Fmax from this workflow compares arms with each other and is never a board number.

Related

• Image finding on the same day: eced1cdd (pinned here, W945) predates nextpnr-xilinx 226609704, which disconnects DSP CARRYCASCIN/MULTSIGNIN tied to GND. On gf16_mul, eced1cdd nextpnr fails $PACKER_GND_NET whatever Yosys made the netlist (cross run 37264809032). The _add_top arms are adders; arms that instantiate a DSP multiplier would hit it. Repinning is the owner's call -- see epic: faster FPGA builds through .t27/t27c and by removing CI waste #851.
• 🎯 EPIC · Matrix [83 formats × {SW-bitexact / decode-HW / compute-HW}] on AX7203 #199 (hardware claims), epic: faster FPGA builds through .t27/t27c and by removing CI waste #851 (build-speed epic).

🤖 Generated with Claude Code