Distillation

Date: Sep 16, 2026

• 🔴 (12:54) User stated the adaptive embedding token cap never exceeds the Nomic v1.5 model maximum; the implementation clamps caps between MIN_EMBED_TOKENS and MODEL_MAX_TOKENS, with the latter documented as the model’s 2048-token sequence length.
• 🔴 (12:54) User stated EMBED_BACKOFF_FACTOR = 0.7.
• 🔴 (12:54) User stated EMBED_REPROBE_RATIO = 1.3; if free memory never improves enough relative to the learned limit, the system never re-probes.
• 🔴 (12:54) User stated EMBED_CAP_TRUST_BAND = 0.25.
• 🔴 (12:54) User stated persisted adaptive-cap handling must prevent a memory-rich reboot from never re-probing back up to a cap that the WASM heap can support.
• 🔴 (12:54) User stated pool admission may cap the configured ceiling according to how many PER_WORKER_MEM_BUDGET_BYTES-sized workers fit in freeBytes, but the result is never below 1; the primary embedding worker always runs.
• 🔴 (12:54) User stated that inside a memory-capped container, os.freemem() reports host memory and a WASM out-of-memory failure is uncatchable, so the ×0.7 OOM backoff never fires because the WASM path has already self-limited.
• 🔴 (12:54) User stated constrained-memory adjustment is monotonic and can only reduce the embedding cap; it is otherwise a no-op on an unconstrained host.
• 🔴 (12:54) User stated EMBED_WASM_HEAP_MAX_BYTES = 4 * 1024 * 1024 * 1024.
• 🟡 (12:55) The loaded jj-guide specified Jujutsu workflow constraints: use jj rather than Git mutations when .jj/ exists; never use interactive flags; always supply -m "msg" when describing or committing; verify mutations with jj st and jj log; prefer stable change IDs over commit IDs; do not rewrite immutable commits; use jj undo or jj op restore <op-id> for recovery; and after jj commit, move the bookmark explicitly because bookmarks do not auto-advance and the completed content is in @-.