mimcap: size the via array off the MIM's enclosure, not the metal's

[carloscl03]

mimcap draws FuseTop at size and hands size straight to via_array, which decides how many via columns fit using the generic metal-over-via enclosure — 0.12um for met5 over via4 on gf180. The MIM plate has a stricter rule: MIMTM.4 (option B) and MIM.4 (option A) both want 0.4um of plate over the via that contacts it.

The plate-to-via margin is therefore a sawtooth in size: it grows until one more column fits, then drops back.

size	via columns	margin
5.00um	5	0.770
5.03um	5	0.785
5.05um	6	0.395
5.80um	6	0.770
5.85um	7	0.395

The floor of that sawtooth is 0.390um: ample for 0.12, and 10nm short of 0.4. So the violations do not scale with the size, they repeat with the via pitch of 0.80um — a hand-picked round size looks clean while a generator that derives the plate side from a target capacitance lands on them regularly.

Shrinking the array region by the difference between the two rules lifts the whole sawtooth by that amount, floor included: 0.390 → 0.670um. The rule goes in the PDK as capmet.via_enclosure; PDKs that do not declare it keep the old sizing untouched.

Cost, stated up front: a column enters later than it used to, so some sizes get one fewer via row and column — 5.05um goes from 6 to 5. That is more series resistance into the plate, which a MIM can afford, in exchange for an enclosure that holds at every size instead of at most of them. A minimal fix (shrink by one grid unit) would keep the vias and leave the margin sitting exactly on the rule; that seemed like the wrong trade.

Verification

Sweeping 5.00–8.00um in 0.01um steps and measuring the FuseTop-to-via4 distance in the GDS, rather than counting violations:

before   301 sizes, 5 below 0.4um (5.05 5.85 6.65 7.44 7.45), worst 0.390
after    301 sizes, 0 below 0.4um,                             worst 0.670

DRC agrees on the nine sizes it was run for: 24 MIMTM.4 at 5.05um and 28 at 5.85um before, 0 everywhere after. A 14-case cell sweep that had 48 MIMTM.4 in one case now passes all of them.

Independent of #102/#104 and of the other open PRs — applies cleanly on main on its own.