Commit d5da57d

- 80a0e88: Make imports that are concurrent with the current version only on map/counter

containers as fast as ordinary linear imports.

Previously any concurrency between the new ops and the current version

(for example a single unmerged head from another peer) forced every later

import to retreat to a critical version and replay the whole branch, and the

map diff scanned every key of every touched map. Per-import cost grew with

both the branch length and the map size: 400 small imports on a 2000-note doc

with one stale head took ~2.4s natively; they now take ~70ms, flat per import

(a linear import of the same chain takes ~50ms).

The oplog now proves, per container, whether the concurrent old history

touches anything that needs positional context. When it only touches registers

the import replays from the current version; the affected maps are still

resolved from history so persisted state that drops tombstones (deleted roots,

dead containers) cannot influence the result. Map diffs in the conservative

path are also restricted to the keys the update touched.

"version": "1.16.1",

- d9ddfba: Make repeated concurrent imports fast: retreat the replay base to the latest

multi-head critical version instead of the single-head one (which is stuck at

the initial fork point on criss-cross sync DAGs), and let the text/list diff

calculators trust a certified-critical base instead of rebuilding their

trackers from the whole history on every concurrent import.

- 80a0e88: Make imports that are concurrent with the current version only on map/counter

containers as fast as ordinary linear imports.

Previously any concurrency between the new ops and the current version

(for example a single unmerged head from another peer) forced every later

import to retreat to a critical version and replay the whole branch, and the

map diff scanned every key of every touched map. Per-import cost grew with

both the branch length and the map size: 400 small imports on a 2000-note doc

with one stale head took ~2.4s natively; they now take ~70ms, flat per import

(a linear import of the same chain takes ~50ms).

The oplog now proves, per container, whether the concurrent old history

touches anything that needs positional context. When it only touches registers

the import replays from the current version; the affected maps are still

resolved from history so persisted state that drops tombstones (deleted roots,

dead containers) cannot influence the result. Map diffs in the conservative

path are also restricted to the keys the update touched.

- d9ddfba: Fix the richtext diff calculator's style table growing a duplicate entry for

every re-replayed style op (e.g. checkout walks after a retreat).

version = "1.16.1"

"version": "1.16.1",