[jeffhsu3]

Branch: placement-helper-ci-coverage → main · 2 commits

two_transistor_interdigitized and common_centroid_ab_ba had no DRC or LVS
coverage. They are also the only consumers of the multiplier's inter-finger
top-metal ports (*sd_top_met_*), so a change to inter_finger_topmet could
break them silently. This adds them to the CI matrix — and the first run found a
real bug.

The bug the coverage found

common_centroid_ab_ba fails gf180 DRC with 4× M2.2a (min metal2 spacing,
0.28 µm), symmetric top and bottom:

M2.2a  x[2.15,2.65] y[-7.205,-6.955]      M2.2a  x[2.10,2.65] y[6.955,7.205]
M2.2a  x[1.975,2.725] y[-7.205,-6.955]    M2.2a  x[1.975,2.725] y[6.955,7.205]

The guard ring was sized off the bbox before any routing existed:

tapref = comcentroid << tapring(pdk, evaluate_bbox(comcentroid, padding=1))

At that point comcentroid holds only the four placed FETs. The routing is added
afterwards and reaches ~1.84 µm above them, well past the 1 µm padding:

	y
device diffusion tops out	5.115
left-side route (0.5 µm tall)	→ 6.155 ✓
right-side route (1.3 µm tall)	→ 6.955
tap ring's met2 rail starts	7.205

0.25 µm where 0.28 is required, short by 0.03. The asymmetry is why it fired on
one side only. Note get_grule("met2")["min_separation"] is 0.3 in glayout's
gf180 rules — stricter than the deck — so nothing was computing too small a
spacing; the routing simply wasn't accounted for.

Fix: place the guard ring after all routing, so evaluate_bbox sees the real
extent. No magic numbers, and it stays correct if the routing changes.

Impact

• common_centroid_ab_ba gf180 DRC: 4 violations → pass; sky130 still passes
• Size: common_centroid_ab_ba grows 1.48 µm in y (15.59 → 17.07, width
  unchanged). Its only consumers, diff_pair and diff_pair_ibias, are
  exactly unchanged at 23.56 × 19.32 and 20.00 × 43.08 — their own outer
  rings already cleared it, and both still pass DRC and LVS on both PDKs.

Coverage caveats

• common_centroid_ab_ba gets DRC only. It sets info["route_genid"] but never
  info['netlist'], so LVS skips it (cells without inputs). Worth adding later.
• two_transistor_interdigitized gets both, and passes both — but its LVS only
  passes with the .subckt fix in the LVS-harness PR. Land that one first,
  or gf180 LVS here reports "inconclusive".

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01RExJUxLyBEHcQRb6GprVSL