v1.7.1

Fixed

• Digest and reflect wrote 1–10 into the 0–1 salience field, and ranking multiplied it straight through.

  detectSalience scored a document 5–7 by design — its comment said "the 1–10 scale used by observe", which observe had not used for a long time — the observe / reflect / predict steps subtracted whole points from that and stored the result, and the reflect tool stored a literal 6. Only the extract step divided by 10. The ranking factor in query and context is 0.5 + salience * 0.5, so a memory minted from one of those observations carried a factor of 3–5 against a ceiling of 1 and outranked every relevant neighbour: on one live store a 7.0 memory with similarity 0.41 sat above a 0.55 match, the L0 context tier (salience × retrievability) put a 9.0 memory first in every session, and 15 of 679 memories (3.0–9.0) plus 70 of 3,222 observations were on the wrong scale — 41 of them unprocessed from the 2026-09-13 journal backfill, waiting for the next dream to mint them. fozikio memory maintain knew the field was 0–1 but repaired by clamping, which would have turned a mundane 5.0 chunk into a 1.0 memory of maximum importance.

  normalizeSalience (new, src/engines/salience.ts) treats any value above 1 as the legacy scale and divides by 10, then clamps to [0, 1]. Digest now scores on 0–1 throughout (0.7 active / journal / mind, 0.6 knowledge or long, 0.5 otherwise; chunks at −0.2, reflect and predict items at −0.1, floor 0.1) and normalises an explicit override, so a frontmatter salience: 8 still means 0.8. reflect stores 0.6. The observe, wonder, speculate and digest tools normalise their salience argument and observe normalises the LLM auto-score. The dream create phase normalises an observation's salience at promotion, so legacy unprocessed observations mint correctly with no store rewrite. query and all three context tiers normalise before computing the factor, so a legacy memory stops hijacking ranking before any repair runs. maintain fix repairs an out-of-range memory by scale instead of by clamp.