Struct Gathered 

pub struct Gathered {
    pub buckets: BTreeMap<&'static str, Vec<ExportRecord>>,
    pub retired: HashSet<String>,
}

Walk the index, bucket every ledger entry by ConclusionClass. Each bucket is sorted by id alone — entry_ids are deterministic hashes (blake3-derived for Plan G thinking, content-derived for others), so id-sort gives:

1. Byte-identical output across runs when the entry set is unchanged (re-export is a no-op).
2. Conflict-resistant git merges: hash-distributed positions mean two developers’ concurrent same-day inserts land in different file regions and don’t textually collide. Plan K t-001.

The previous sort key was (created_at, id). That gave the same byte-stability but clustered same-day inserts together, raising the odds of merge conflicts when two devs added entries on the same day. id-only loses the time-ordered visual diff but gains real conflict-resistance — the Plan K principle (sidecar = judgment; conflicts must be meaningful) prefers the latter.

Upgrade note: re-exporting an existing project at >=1.0.36 will produce a one-time mass reorder of every .asd/conclusions/*.jsonl file. Commit it once, then steady state. What a gather produced, plus what it deliberately withheld.

Fields

buckets: BTreeMap<&'static str, Vec<ExportRecord>>

retired: HashSet<String>

Entry ids excluded on purpose — superseded, or a hypothesis below the confidence floor. Everything else missing from buckets is missing because this clone could not produce it, not because it decided not to.