Documents how each backup path has actually been exercised — Google
Auto Backup verified on emulator (status card, blob round-trip via local
transport, restore card by screenshot; decrypt+import by instrumented
test since ADB can't tap the Compose dialog reliably), file import/
export verified earlier, S3 still pending a live bucket. Captures the
scrypt wf=15 rationale and the safe real-hardware test protocol.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Amends decision #8: the producer now lives in tools/fest_flatten.py
(stdlib Python, streaming iterparse over the 115 MB M30 XML) instead of
an out-of-repo server job. It pulls fest251.zip from dmp.no, flattens
8934 active human-use brand entries (name/strength/unit/form/ATC +
modal package size from pakningsinfo), and writes the asset checked in
at app/src/main/assets/fest/slim.json — so autocomplete works offline
out of the box. preReleaseBuild depends on refreshFestData: fast-exits
under 30 days, keeps the old file on download failure (offline release
builds never break). FestRepository: downloaded cache wins over the
bundled asset.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Milestone 8. docs/fest-dataset.md is the interface to the out-of-repo
server job that flattens the M30 Rekvirent extract (FEST, not
Felleskatalogen — decision #8) into slim JSON. The repository
downloads on demand, parses BEFORE overwriting the cache (a failed
refresh can never clobber a working dataset), and serves all searches
from memory. Ranking (prefix > word-prefix > substring) is a pure
object with tests. Every failure mode degrades to manual entry —
autocomplete is a convenience, never a gate.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
