Distillation

Date: Sep 16, 2026

• 🟡 (13:18) Focused validation passed all 250 tests across 6 test files in Vitest v4.1.8; duration was 5.52s (transform 4.39s, setup 4.03s, import 2.19s, tests 5.29s, environment 1ms). Coverage included the three embedding-pool admission blockers and native fallback.
• 🟡 (13:18) Assistant noted that passing tests alone do not prove the cgroup scope and reservation guards, and began auditing the corrected candidate for edge cases before locking mutation evidence.
• 🔴 (13:18) User-provided Jujutsu workflow states that jj never fails on conflict: rebase, new, and squash record conflicts in the resulting commit, after which jj st must be used to identify conflicted files and they must be resolved manually rather than with interactive jj resolve.
• 🔴 (13:18) User-provided Jujutsu workflow requires using jj rather than git for mutations whenever .jj/ exists; forbids interactive flags; requires -m "msg" for describe/commit operations; and requires verifying squash, abandon, rebase, restore, and commit mutations with jj st and jj log.
• 🟡 (13:18) Current ordered work plan: 1. “Finish cgroup discovery and inherited-boundary regressions and fix” — in_progress, high priority; 2. “Fail closed primary embedding sizing when constrained headroom is unknown or conflicting” — pending, high; 3. “Reserve unreflected memory during burst embedding-pool growth” — pending, high; 4. “Complete native runtime resolver and one-shot WASM fallback proof” — pending, high; 5. “Prove critical guards with base failures and mutation checks” — pending, high; 6. “Run full validation and freeze exact Jujutsu candidate” — pending, high; 7. “Obtain independent correctness and security reviews” — pending, high; 8. “Commit completed Jujutsu change and report runtime guidance” — pending, medium.
• 🔴 (13:19) User-provided implementation rationale states that native ONNX container OOMs can trigger an uncatchable cgroup OOM-killer SIGKILL, so the ×0.7 OOM backoff never fires; the WASM path instead self-limits against its fixed 4 GiB heap.
• 🟡 (13:19) packages/core/src/embedding-cap.ts documents why EMBED_TOKEN_CEILING remains shared between WASM and native paths: WASM is bound by its fixed 4 GiB heap, while native uses the same ceiling for memory prudence and quality because nomic-embed-v1.5 effectiveness plateaus around ~2048 tokens, the approximately 4962-token ceiling sacrifices negligible quality, and long inputs are chunked upstream.
• 🟡 (13:19) Assistant assessed that packages/core/src/cgroup-memory.ts already contains the audit’s hierarchy traversal, mount selection, namespace-root handling, and live cgroup-move corrections; next steps were checking each regression against real production call paths, reducing any overbroad code, and then mutation testing.