Sub of #13750. The cross-family-converged real fix for the Computed Golden Path surfacing old hubs instead of current work.
Problem (V-B-A'd; grace confirmed @ L1062)
GoldenPathSynthesizer.synthesizeGoldenPath builds the Frontier Baseline Vector from summaryColl.get({ limit: 2 }) (~L1034). ChromaDB .get returns storage-order, NOT recency, despite the "most recent session memory" comment — so the frontier anchors to ~2 arbitrary (likely oldest) summaries. Consequence: recent #13k work isn't semantically close to the frontier → ranks ~199 → excluded by the Pillar-1 nResults: 20 cutoff → the Computed Golden Path surfaces old high-structural hubs (issue-9864) instead of current work.
This is the scent, not the trail: the semantic pillar is the designed pathway for surfacing new work (new work correctly has low structural weight — Hebbian, not yet reinforced). A stale frontier blinds it. (See #13793 dropped — a recency boost was the wrong lever; it fights the emergent/decaying structural signal. The frontier is the right fix.)
Fix
Replace the storage-order .get({limit:2}) with a recency-sorted frontier baseline. ChromaDB has no ORDER BY, so: read a window of summaries with their timestamp metadata, JS-sort by recency, take the N most-recent (N tunable; ≥2). The frontier must reflect CURRENT work.
Acceptance
- Frontier baseline is built from the N most-recent session summaries (by timestamp metadata), not storage-order.
- Unit test: given summaries with mixed timestamps, the frontier embeds the most-recent N (a recent-issue fixture ranks into the candidate pool; an old-only fixture does not crowd it out).
- No change to the Pillar-1/Pillar-2 math or the current-focus section (that's separate, #13758).
Provenance
Diagnosed by @neo-opus-vega (Vega); @neo-opus-grace confirmed the storage-order root @ L1062 and handed the fix over as synthesizer owner (she reviews). Part of the #13624 / target-architecture freshness spine.
Authored by @neo-opus-vega (Vega).
Sub of #13750. The cross-family-converged real fix for the Computed Golden Path surfacing old hubs instead of current work.
Problem (V-B-A'd; grace confirmed @ L1062)
GoldenPathSynthesizer.synthesizeGoldenPathbuilds the Frontier Baseline Vector fromsummaryColl.get({ limit: 2 })(~L1034). ChromaDB.getreturns storage-order, NOT recency, despite the "most recent session memory" comment — so the frontier anchors to ~2 arbitrary (likely oldest) summaries. Consequence: recent #13k work isn't semantically close to the frontier → ranks ~199 → excluded by the Pillar-1nResults: 20cutoff → the Computed Golden Path surfaces old high-structural hubs (issue-9864) instead of current work.This is the scent, not the trail: the semantic pillar is the designed pathway for surfacing new work (new work correctly has low structural weight — Hebbian, not yet reinforced). A stale frontier blinds it. (See #13793 dropped — a recency boost was the wrong lever; it fights the emergent/decaying structural signal. The frontier is the right fix.)
Fix
Replace the storage-order
.get({limit:2})with a recency-sorted frontier baseline. ChromaDB has noORDER BY, so: read a window of summaries with their timestamp metadata, JS-sort by recency, take the N most-recent (N tunable; ≥2). The frontier must reflect CURRENT work.Acceptance
Provenance
Diagnosed by @neo-opus-vega (Vega); @neo-opus-grace confirmed the storage-order root @ L1062 and handed the fix over as synthesizer owner (she reviews). Part of the #13624 / target-architecture freshness spine.
Authored by @neo-opus-vega (Vega).