riemann

v1.53 — NS-46: fine translated-radical blocks

Classification: analytic proofs and an explicit open input. This is not a numerical illustration or a certified computation. No new window or tail metric.

Result and scope

The construction changes the v1.35 block: centers remain in [-1,1] while their spacing shrinks, and a finite interpolation estimate explicitly pays the Gram conditioning cost. For a=log(lambda) sufficiently large, set J=floor(lambda²/(40000a)), m=2J+1. The complete cutoff block has rank m, exact even/odd ranks J+1 and J, and complete operator residual at most C exp(-lambda²/2000). Thus m is asymptotic to lambda²/(20000 log lambda).

Removing o(lambda²/log lambda) directions cannot leave a uniform positive gap or a positive gap bounded below by a fixed power of lambda^(-1). Smaller positive gaps are not excluded. This is an obstruction to that positive coercivity objective; it does not exhibit a negative Weil direction or bound the remaining complement. The actual even source removes at most one even block dimension. Full prime and pole terms, and all block mixtures, are retained.

The complete, signed, uniform complementary lower floor remains an open input. Its transfer to the full form was already proved in v1.18/v1.36. No new transfer lemma is claimed. Neither G2 nor RH is proved. The constants and eventual starting threshold are analytic; no finite numerical threshold is certified.

Files and verification

Replay the actual build with ./manuscript/build.sh; validate this incremental archive with python3 tools/verify_manifest.py v153. Check the external inventory by comparing each path, byte length and SHA256 in provenance.json. This integrity check does not replace analytic review.

Local draft based on commit 24f4ab5 (PR #21), with standalone task claim 0355edb. No v1.53 publication, merge or tag is asserted. Main advanced to b0baadf while work was isolated; the PR #21 dependency is retained explicitly.