fix(verify): a missing GPU core count is a gap, not a contradiction

[Seungpyo1007]

Problem

vendor_core_field fails whenever cuda_cores (nvidia) or stream_processors (amd/intel) is absent. Both are already in RICH_FIELDS["gpu"], so completeness scores that same absence — charging it again under consistency bills one gap twice, under a name that says the record disagrees with itself.

Splitting the 281 flagged records by what they actually contain:

	records
neither field set — a gap	279
the count filed under the other vendor's field — a contradiction	2

The gaps are not even all gaps. They include the NV1, RIVA 128, RIVA TNT and GeForce 256 — parts that predate the unified shader, where a CUDA core count is inapplicable, not missing.

Change

• own vendor's field present → pass
• other vendor's field present while its own is absent → fail (a real mismatch; still soft)
• neither → na

Three regression tests cover the three branches; tests/verify is 56 passed.

Effect

All 281 flagged records are yellow with consistency at exactly 23.3/35 (= 35 × 2/3 — one failure out of three evaluated signals). Dropping the false failure returns those ~11.7 points. 147 of them are in the near-miss band (T1/T2 host, score 65–74) and should cross the green threshold of 75; the rest are held back by other things.

Found while looking for what remains between the near-miss pool and green: of 2,907 near-miss records, 1,926 need only 2 points. For smartphones that pool turned out to be unreachable — the missing rich fields (charging_wired_w 100%, os_version 100%, cameras 100%) have no in-record source, and every one of the 2,385 records lacking os_version carries a bare os string with no version in it at all (Android, Windows Mobile, Feature phone, Symbian). The GPU pool is different: nothing needs to be sourced, the scoring was just wrong.

Refs #1