Instead of fetching up to 50k inventory snapshots (~442MB), scope the
query to only the time range of the current submission batch with a
15-minute buffer. For a typical 1-hour batch this drops from 50k rows
to ~300 rows, well within the 1280Mi pod memory limit.
The coalesced_at_idx on intent_inventory_snapshots covers the BETWEEN
clause so this remains efficient.
Proof: history-writer OOM kills from refreshQuoteOutcomes inventory fetch
Assumptions: 15min buffer covers the attribution window for all submissions
Still fake: heuristic gap outcomes may attribute trades imprecisely
