[drakestapleton]

The same immutable server delivered the first visible response for 32,768 input tokens in 104.728 seconds with cached history versus 108.372 seconds before; fresh follow-up was 103.074 versus 108.918 seconds. Full-model reading measured 109.500 versus 103.451 seconds with identical scores. These are single shared-machine observations; repeated 5,017-token total was essentially unchanged.

This carries the 30B cached-history mechanism into 27B attention. Omega validates requests and directs allocation, preparation, completed full-history copy and attention through one reusable buffer. Original history and arithmetic remain authoritative. The per-call API has no global selector: mode 0 preserves ordinary behavior, mode 1 mirrors supported calls, and mode 2 applies a provisional 1,024-row threshold. Failures cannot advance to output publication.

Completed verification includes the complete 83-target native host suite, full Core preflight, 63 physical attention checks, six complete-score cases, seven threshold/tail cases, long32 plus short-after-long, and deliberate corruption controls. Every compared score bit matches. Short serving covers streaming, seeded sampling, seven coding cases and prose; paired32K serving matches answers and token IDs across settings. The 65,536-token candidate run passes retrieval and saved/fresh continuation parity. Its first visible response was 237.355 seconds, saved follow-up 0.531 seconds and fresh follow-up 237.970 seconds. This candidate-only 64K run is a correctness observation, not a paired speed claim.

The host-suite registration defect is fixed: carried-tests.mk appends instead of replacing targets registered by earlier fragments. The actual graph retains all 81 previous targets plus both new attention targets. The full 83-target run exits 0 at 9ade9f5; production source and the tested binary remain unchanged. Earlier failures and historical receipts remain preserved.

Integration evidence and host83/64K follow-up pin sources, binaries and original logs. The 128,000-token serving test now passes retrieval and saved/fresh continuation parity; first visible response was 570.312 seconds, saved follow-up 0.778 seconds and fresh follow-up 570.967 seconds. This is a candidate-only correctness run, not a paired speed claim. Five alternating quiet timing pairs and independent review remain open; no quiet qualification or default promotion is claimed.

Stacked on #259; tracked by aien-dev/aien-sovereign-core-mv#710. Native/stub compatibility, API 0.13, pinned Bridge regeneration and Crumb checks pass. No new kernel, production dependency, default change or deployment. Apache-2.0 WITH LLVM-exception notices retained.

Completed 128K evidence preserves the immutable server/source identity and original event hashes. All planned short/32K/64K/128K serving correctness gates now pass; quiet alternating timing and independent review remain open.

[aien-spark]

Merge desk note (AI), 2026-11-01.

Independent review of the whole stack (#251, #252, #257, #258, #259, #260): APPROVE-WITH-NOTES, no blockers. The stack merges cleanly onto main. The #246 GB10 admission files are byte-identical on this head and on main, so there is no duplicate rule.

Default-path changes (switches off), each output-equivalent or a hardening:

1. omega_gpu_cmdlist.c (27B: exact GPU head and reusable command seals #251): every command list now routes launch seals through omega_native_mcreseal. It uses the same SHA-256 over the same normalized bytes as mcc_seal, plus a byte-compare memo, and every other gate still runs.
2. omega_blackwell_codegen_qg.c (27B: test an Omega 32-row prompt kernel #252): the store-offset bound is now the exact immediate limit from admit.om. Default rt=4 F32 widths from 32768 to 37449 were refused and are now admitted. The 27B uses widths of at most 17408, so it is unaffected.
3. omega_gpu_prefill_attn256_api.c (perf(attention): add scoped Omega-controlled cached history #260): a failed session_on_close registration now returns CHIP_FAIL. It used to be ignored.

Hashes check out: for all nine units, the src= and c= values in cudapp_om_units.ident match the .om parts and the committed C. A byte-for-byte regeneration with units.sh --check needs the pinned omc and was not rerun in this review.

Follow-up, not in this stack: set_exact_head copied its coefficient table to the chip without checking it, so a NaN, negative or infinite entry could pick a wrong greedy token. A fix with a host test and a mutant is on branch fix/q38-exact-coef-check and will be opened against main after this merges.

[cursor]

Risk: high. I left a non-blocking comment and did not approve: this stacked attention/runtime/nvrm change is above the medium approval threshold. Cursor Security Agent completed with no findings, Cursor Bugbot was not present, and aien-spark was assigned for human review.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}