[ruvnet]

Saving and reopening an INT8 TypeSafe decision system changed five of 150frozen ticket decisions. Native ONNX inference now uses one text per tensor, and its identity includes #ort-singleton-v1 so frozen-head consumers can distinguish old batched features. The regression fails before the fix and passes afterward.

This maintenance PR also fixes the latest main-branch failures: ruvllm serve binds before announcing readiness; Ubuntu CI uses a checksum-pinned compatible ONNX runtime; graph rebuild correctness is tested with synchronized concurrent access instead of a scheduler-sensitive max-latency assertion; nine compatible npm lockfile updates clear the production audit.

Validated candidate: 10bf19e1c055d397de8bec8dab15e80f716e7a51.

• Entire published reproduction script passed: build/typecheck, 19 embedding unit tests, exact INT8 regression, 7 server tests, 10 graph correctness tests, production audit, 33 benchmark tests, and six fresh-process RVF/HTTP runs.
• FP32 and INT8 each match 900/900 direct-versus-restored records (150 distinct tickets, three runs, first and warm passes). Corrupted RVFs are rejected.
• Final shared-runtime medians: FP32 warm p95=14.23ms and cold start=6.23s; INT8 warm p95=7.68ms and cold start=3.37s. INT8 RVF is 34,873,088bytes versus 133,952,128bytes. These are local host measurements, not a controlled remote-JEV speed comparison.
• Shared-runtime INT8 department/urgent/frustration accuracy: 84.7%/81.3%/78.0%; ECE 0.04982. Static-runtime results differ and are preserved separately. Runtime changes still require qualification.
• Full JavaScript suite: 330 passed, 0 failed, 1 WASM skip, with the shared runtime in LD_LIBRARY_PATH. Exact-candidate remote CI is pending; no complete release qualification is claimed.

Detailed report, all receipts and validated reproduction script. Unpublished source release draft, created by GitHub Actions; downloaded source archive matches the exact candidate bytes (SHA256 84b5c30b1d4faa0b1e8d94c04ed2826f052b887175f85787ca61ec792dacd9a5). No npm/crates publish or merge has occurred.

Tracks #1177, #1178, #1179; remaining qualification is explicit in #1181. Benchmark scope is TypeSafe sentence embeddings with frozen JEV replay, not KGE. Existing workspace peer/platform conflicts and other manifests' Dependabot alerts remain outside the production-audit repair.

[proffesor-for-testing]

Independent October21 review of the graph-only two-file patch at10bf19e1c055d397de8bec8dab15e80f716e7a51, Codex acting for Dragan.

Reproduced old timing test on main46d816ddc: max290.480ms versus200ms. Three CPU-affinity repeats passed. Instrumented controlled contention reproduced310.958ms comprising15.007ms graph-lock wait,253.522ms CSR rebuild and42.426ms search; another291.592ms read waited only1.38microseconds for graph lock. The old oracle measures several stages and scheduling, rather than isolating off-lock behavior. This does not establish the exact hosted CI cause.

Applied only this PR's existing graph helper/test patch to main:10tests pass,1existing ignored, twice. Mutation control holding the graph write lock across the builder is rejected immediately in0.02s. The deterministic replacement is therefore useful evidence for off-lock construction and exact mutation replay. No duplicate patch is being submitted; the existing ignored latency benchmark remains available for numeric performance qualification.

This review does not cover the PR's other maintenance changes or resolve its current conflicts/failing checks.