#768: the Sylow divisor condition — Eric Li's constant, independently rebuilt, now the THIRD clean Lean check

The target (erdosproblems.com/768). Let A be the integers n where every prime p∣n has some divisor d∣n, d>1, with d≡1 (mod p) — the "Sylow divisor condition." Erdős showed |A∩[1,N]|/N = exp(−(c+o(1))√(logN·loglogN)) for some c>0 bounded between the trivial c=1 and c can't-be-improved; the question is the exact constant.

Already flagged, already double-checked — we're a third opinion, not a discovery. Eric Li posted this himself on 12 Jul, with a Lean formalization and full disclosure of AI assistance. Two days later, a different person (JohanLand) independently formalized the same paper from scratch — without having seen Li's own Lean code — and it also compiled clean, catching one trivial, non-load-bearing gap in the paper's Lemma 6.3 that doesn't affect the result. That's real, working peer scrutiny already in public. What we added: a third independent build, on infrastructure neither of them touched.

What we did. Cloned github.com/ericlisg/erdos768-lean fresh, built the whole ~8,000-line project from source (including its real analytic input, MediumPNT from PrimeNumberTheoremAnd — a real, substantial prime-number-theorem dependency, not a stub), then ran our own #print axioms:

'Erdos768.erdos_768' depends on axioms:
[propext, Classical.choice, Quot.sound]

Clean. The repo's own build already enforces this via a hard-coded #guard_msgs check that fails the build on any axiom mismatch — we independently confirmed the same result outside that guard.

Constant: c = 1/(2√log2) ≈ 0.849, resolving the open part of #768.