S89: hold before replacing the typing editor¶
Scope: issue #179, following the native n14 failed run. Its main note held
ABCDIJKLMNOPQRST, its existing copy held ABCDEFGH, and a later Resume
made a second copy. That run remains a failure: the issue requires all twenty
characters together in the main note after Sync now, not merely a union of
files. No timestamp field or note body is discarded to meet the copy bound.
Repair¶
A device can receive the other device's authenticated answer: true branch
before the background author sees its competing edit. When the attempted
merge overlaps and this device has recent trusted editor input, it now holds
before the ordinary conflict rule can replace the editor's saved text. Clean
merges still complete, and two actual typists without the background flag do
not take this path. The editor-side role is persisted and the existing v3
control is published. The recipient uses its own current answer proof for the
exact target file to retain the background-author Resume role. A same-name
local file of another identity cannot supply that proof. The control still
deduplicates; the wire format and server API do not change.
Background Resume first durably advances the proven copy to its latest held bytes, retaining the baseline in that copy's history. It can then adopt the single current peer branch instead of running the ordinary fork rule and making a second copy of that peer. Multiple, missing and local-author heads are left for ordinary reconciliation. Moved, deleted and multi-chunk peer heads are refused. A save during the peer download leaves the main note and hold intact. Existing independent-edit and split-copy refusals remain in force.
Reproduction and checks¶
Pinned Node 26.8.2 and TypeScript 5.9.3. The directed fixture publishes real
encrypted A/B siblings in the native order: common ABCD, A's typed EFGH,
then B's automatic stamp of the older ABCD. Before the fix, both fixture
orders made the first copy instead of holding. It then runs both engines,
both stampers and both open editors, appends IJKLMNOPQRST, and lets the
stampers run for five virtual minutes while paused. After one Sync now per
device, in both orders, both main notes hold ABCDEFGHIJKLMNOPQRST, exactly
one copy holds B's latest paused snapshot, both engines are idle with no
pending or in-flight work, and another quiet interval adds no journal frames.
The fixture controls delivery order before starting the hold interval; native
Obsidian validation is separate and is not claimed by this receipt.
- Clean full plugin suite: 1,100/1,100, zero failures/skips,
112071.651083 ms (
179-overlap-restored-full.log). This includes #135 co-typing, both editor states, both existing and new Resume orders, same-size rewrites, copy-edit refusals, and the legacy decoder compatibility checks. - New focused guards: 18/18 after restoring all mutations
(
179-overlap-final-guards.log). - M690–M709: 20 compiled; all 20 killed by 36 focused test failures.
- Recut M413, M480, M496, M533, M535, M560, M563, M568–M571, M579 and M620: 13 compiled; all 13 killed by 34 focused failures. The recuts change context, not the old hostile substitutions. M413's first focused filter missed its overlapping branch; a direct unpushed same-line regression now kills it. No timing collateral is counted.
- All 654 patches apply with
patch -F0 --dry-run; no fuzzy application. - Raw logs and final
summary.jsonare in local evidence179-overlap-mutants/; earlier focused results are retained infirst-run/. Focused counts are not full-suite mutation counts. The coordinator owns the final complete matrix and native desktop/phone acceptance.
New mutation results¶
| Mutant | Target | Focused failures |
|---|---|---|
| M690 | the typing peer makes a conflict copy before holding | 3 |
| M691 | ordinary user edits are treated as automatic answers | 2 |
| M692 | a passive editor originates the typing-side hold | 1 |
| M693 | the active editor resumes as the background rewriter | 3 |
| M694 | the editor-side hold is not persisted | 1 |
| M695 | the editor-side hold is not shared | 1 |
| M696 | the typing-side hold still falls through to conflict copying | 2 |
| M697 | parallel typing-side detection emits duplicate notices | 1 |
| M698 | the received control classifies its own id instead of the note | 1 |
| M699 | the answer author resumes as a remote editor | 1 |
| M700 | background Resume restores the stale local branch before reconciling | 5 |
| M701 | background Resume guesses one branch from several peer heads | 1 |
| M702 | a local in-flight head is mistaken for the peer editor | 1 |
| M703 | Resume selects an unrelated historical version instead of the peer head | 5 |
| M704 | Resume records the peer head over restored local bytes | 2 |
| M705 | Resume records the old version over the peer bytes | 1 |
| M706 | Resume writes the peer under a different path | 1 |
| M707 | Resume admits a deleted or oversized peer head | 2 |
| M708 | the peer answer pauses even a clean nonoverlapping merge | 1 |
| M709 | the control claims another file at the same path is its author | 1 |
The one-chunk shape guard also refuses deletion: the validated manifest contract gives deleted versions zero chunks. It does not need a second, redundant deleted-field predicate.
Native rerun and remaining acceptance boundary¶
The coordinator's fresh n15 native rerun passed on the final bundle: both main notes retained all twenty characters, exactly one copy remained after one Sync now per device (B first), and five-minute hold and post-Resume observations each added zero note-version POSTs. The desktop receipt records the bundle and matching file hashes. The iPhone S89 result remains unproven. This change prevents the first split; it does not infer how to recombine the old n14 conflict history. The retained n14 failure and its screenshots remain useful evidence. No native app, vault, coordinator source, version lock, reviewer or watcher was changed in the executor lane.