Skip to content

Historical catch-up and typing repairs, 2026-09-24

This is author-operated diagnostic evidence following the native 1.1.3 phone campaign. It distinguishes simulated regressions from native acceptance; it is not an independent review or a final phone result.

A joining device recreated resolved conflicts

Fresh phone pairing on bundle 530ca4aeb1fee893817f99110346226e7573714c542f3d303b45971e0f4b85aa created seven new conflict files from an old typing note. Its journal held 44 versions and one current head; the new copies' creation records followed the pairing. No historical version of that note had been garbage-collected.

The real file endpoint returns its newest ten versions plus every current head. The fake server had returned the entire graph. With the bounded view, the client could not walk old parent links to prove that earlier forks were already resolved and mistakenly preserved them again as new conflicts.

The repair skips an obsolete feed entry when the current head list is nonempty and every head has a version record in the response. It leaves local bytes untouched for later feed entries to advance. An older live head outside the ten-version window still causes both edits to be preserved. Empty and incomplete head views retain the conservative conflict path.

Five regressions in history-catchup.test.mjs cover fresh desktop and mobile replay, an old live head, an empty head view and a missing head record. The two fresh-device cases each made eight spurious copies before repair and none after. M716, M729, M730 and M731 apply, compile and fail 2, 1, 1 and 1 focused tests; the restored source passes. An initial scratch build missing vendored Obsidian declarations is retained as a setup failure, never counted as a kill.

Adjacent lines and same-line appends

The native timing build passed interleaved typing in sections separated by an unchanged line and both rename directions. An adjacent-line attempt produced several copies. Issue #135's acceptance criteria require the exercised typing scenarios to finish with the same note and at most one copy, so the narrower successful native case did not complete that issue.

The old merge intersected the lines unchanged on both sides. Two adjacent line edits had no common anchor between them and were grouped into one false overlap. The repair compares each side's changed base intervals independently. Neighboring replacements, insertions and deletions retain their actual edits; intersecting replacements still refuse to merge.

The same-line append simulation exposed a separate gap in the previous test: it allowed one copy per fork, resulting in five copies in a single typing session. The strengthened test requires the shared main note to contain every unique keystroke exactly once, with no copies. When both versions only append to one existing line, obsync now keeps their shared appended prefix once and joins the remaining additions in lexicographic order. Unicode code points define the shared prefix, so different emoji cannot share half a surrogate pair. Replacements of existing characters and multi-line overlaps remain conflicts.

Before repair, the new adjacent-line desktop/mobile simulation and three primitive witnesses failed. The two append primitives and strengthened same-line session also failed. Focused passing results after repair are recorded separately from the complete gate and native follow-up. Controls M732–M764 each apply, compile and fail their relevant primitive tests; restored source passes all those tests again.

Two existing witnesses were updated to keep their original safety purpose:

  • The write-record race now requires the local replacement and the peer's independently added line in one main note, with no copy. Requiring the replaced common-ancestor text to survive would be requiring an extra copy of text the user deliberately changed. M23 still compiles and fails it.
  • The recursive-base fixture prepends each new word instead of appending it, so a direct append merge cannot bypass the recursive path under test. Three levels still merge and the fourth is refused. M305 and M306 each compile and fail this witness; the original depth limit is unchanged.

Reproduce the targeted suites with the pinned Node toolchain:

npm --prefix plugin run build
node --test plugin/test/history-catchup.test.mjs plugin/test/conflict.test.mjs \
  plugin/test/cotyping.test.mjs plugin/test/merge.test.mjs plugin/test/samename.test.mjs

The pinned full make check passes 1,138 plugin tests, 144 core and 375 server tests, two CLI tests, 70 dashboard tests, 767 contract tests, 94.73% Rust line coverage and both secret scans. All 709 mutation patches apply with zero fuzz. The built plugin SHA-256 is 3e8d499e55ec3b2573bb788826e94a57979fd3be1dfa23a56f5fabec22f2cb47. Earlier full-gate failures are retained: the first sandbox run could not bind test sockets; later failures exposed the stale-head, saved-write and recursive fixture expectations discussed above. None is counted as a passing run.

The complete mutation matrix and real-phone fresh pairing and typing on the repaired bundle remain separate gates. Earlier phone screenshots validate only the bundle and actions named in their original record.