Distillation

Date: Sep 16, 2026

• 🔴 (15:06) User instructed that if recall is needed, use lore recall or http://127.0.0.1:3207/ui/search because the recall tool is currently broken.
• 🟡 (15:06) Assistant reported that the focused memory-admission regressions were green and planned to continue with native-runtime resolver integrity, strict error classification, public diagnostics, bundle artifacts, and mutation-testing each safety guard.
• 🟡 (15:07) Working copy tmzmtlss at commit e6d0f1edc850 (fix/embedding-runtime-admission, fix(core): correct embedding runtime selection and admission) contained 18 changed files totaling 1,818 insertions and 90 deletions relative to parent 2cbf40060c67; packages/core/src/cgroup-memory.ts and packages/core/test/cgroup-memory.test.ts were newly added, while embedding runtime, native ORT, gateway bundle, platform-package, and related test files were modified.
• 🟡 (15:07) Jujutsu reported a bookmark conflict for fix/responses-private-references, with suggested inspection via jj bookmark list and resolution via jj bookmark set <name> -r <rev>.
• 🔴 (15:07) User requires the adaptive embedding token cap never to exceed the Nomic v1.5 maximum sequence length.
• 🔴 (15:07) User requires that when free memory never improves after a real hardware limit is learned, embedding admission must never re-probe.
• 🔴 (15:07) User requires a learned knownBadCap to remain an upper bound across restarts, so a memory-rich reboot never re-probes above a cap already rejected by the WASM heap.
• 🔴 (15:07) User requires worker affordability calculations never to return fewer than 1 worker because the primary worker always runs.
• 🔴 (15:07) User stated that in a constrained environment the WASM path self-limits before an uncatchable process OOM, so the ×0.7 OOM backoff never fires; all freemem-derived decisions must therefore be capped by memory the container can actually provide.
• 🟡 (15:08) packages/core/src/ort-native.ts resolves native ONNX Runtime from per-platform packages named exactly @loreai/onnxruntime-<os>-<arch> for published gateways, or from the gateway workspace’s direct onnxruntime-node dependency; dist-only and unsupported-platform installations return null and use WASM.
• 🟡 (15:08) supportsNativeOrt() allows non-Linux runtimes, but Linux requires a parseable glibcVersionRuntime of at least glibc 2.27; missing, malformed, or unavailable Linux libc information fails closed to WASM.
• 🟡 (15:08) validNativeBinding() canonicalizes both the package directory and binding path with realpathSync(), requires the binding to remain within the package directory and be a regular file, then validates its first 4 bytes: ELF 7f 45 4c 46 on Linux, either cf fa ed fe or fe ed fa cf Mach-O magic on Darwin, and 4d 5a (MZ) on Windows. File-open, read, stat, canonicalization, and close failures return null.
• 🟡 (15:08) Published native-package resolution requires the gateway manifest name to be @loreai/gateway; the platform package must be listed in optionalDependencies; the gateway version, declared platform-package version, and published package version must all match; loreOrtNodeVersion must equal the bundled expected ORT version; and package os/cpu must each be a one-element array matching the runtime.
• 🟡 (15:08) Workspace native-package resolution accepts onnxruntime-node only when declared in the gateway’s dependencies, devDependencies, or optionalDependencies at exactly the bundled expected version; the resolved package manifest must also have name onnxruntime-node and that exact version.
• 🟡 (15:08) Workspace binding discovery canonicalizes onnxruntime-node/bin, requires it to remain inside the resolved package, requires exactly one directory matching /^napi-v\d+$/, and checks <bin>/<napi-dir>/<platform>/<arch>/onnxruntime_binding.node.
• 🟡 (15:08) nativeIntraOpThreads(parallelism = availableParallelism(), logicalCpus = cpus().length) limits native ORT intra-op threads only when cgroup-aware available parallelism is lower than the host logical CPU count. It floors valid values to integers and at least 1, returns the restricted count when avail < total, and otherwise returns undefined to preserve ORT’s physical-core-based default and avoid raising it on hyper-threaded hosts.