Skip to content

Co-typing publication order follow-up, 2026-09-24

The pristine 1.1.3 baseline at cba9eec failed one of six full-suite lanes: part of the peer's typed line remained in a conflict copy. Five lanes passed. The final mutation run was paused; none of these failures counts as a mutation kill. The physical phone run at that source remains evidence only for its recorded input cadence.

Reproductions and repair

Synthetic two-editor traces with delayed platform-crypto completion exposed several related paths through the same failure:

  • An editor upload could read a merge's newly written bytes before the merge receipt updated its parent. Merge writes now reserve the existing per-note publication queue before writing and retain it through the receipt.
  • An upload could finish while a merge prepared its write, leaving no pending promise but a newer local parent. The merge now rechecks the version and retries from the current graph. The deterministic witness previously left two server heads even though every character was present on both devices.
  • Obsolete feed entries consumed the merge-loop budget during catch-up. An incoming version with a retained descendant head is now skipped before counting. A missing graph entry is not treated as proof of supersession.
  • One typist could stop while the peer continued its independent branch. Progress descending from the previous peer version, without incorporating this device's recorded output, starts a new count. Repeating one head, unrelated forks, and replies incorporating this device's output still reach the existing limit. Neither the five-resolution limit nor the three-level criss-cross bound increased.

A previous delayed-receipt regression held its receipt while awaiting another merge. That schedule now waits behind the publication reservation. Its fixture now directly installs the newer durable record while withholding the receipt, retaining the original no-overwrite assertions and the M715 mutation witness. The blocked exploratory run is retained as a fixture failure, not a pass.

Evidence

Eight new controls apply and compile. Their focused witnesses are M716 (one), M717 (two), M718 (one), M719 (one), M720 (two), M721 (one), M722 (one), and M723 (one). The retained M715 also compiles and kills its delayed-record regression. M151, M162 and M714 are mechanically re-cut to the same behavioral changes after the surrounding code moved. All 668 current patches apply with zero fuzz. The first M717 draft did not compile; its corrected compiling run is the mutation evidence, and the compiler failure is not a kill.

Six isolated synthetic co-typing schedules, with crypto-completion delays of 5, 7, 10, 0, 5 and 7 milliseconds, all converge on one server head, identical main notes containing both complete sentinel sequences and zero conflict copies. Their local receipts are cotyping-trace-94 through -99. Earlier failed schedules remain alongside them. These synthetic scheduling controls are not measurements of real network latency.

The candidate bundle SHA-256 is 95b13f24ade8b4760a21216df10ef0cd9fb7b99c139e0b2e65cdd56e516e2e68. The final pinned make check passes 1,114 plugin tests, Rust 144 core + 375 server + two CLI tests, 70 dashboard tests, 767 contracts, 94.73% line coverage and both secret scans. One core benchmark remains intentionally ignored. Six pristine matrix baselines, complete mutation measurement and affected native follow-up remain pending. The existing matrix is historical until its replacement is measured. No Ready or release claim follows from these focused results.

Final baseline follow-up

The subsequent six-lane baseline did not establish a clean final candidate. One lane observed a case-rename fixture before its delayed note moves had finished. Waiting for the folder alone was insufficient: its two note records must match their latest server frames on both devices. A 400 ms transport delay reproduces the old assertion failure; the corrected barrier retains the two-version, identity, path and quiet-scan assertions.

A later baseline exposed a real pending-change race. A note can move before Obsidian reports its rename, while an earlier change is still settling. The missing old path went directly to deletion. It now enters the existing bounded move-discovery path; an old path whose identity was already removed by a pulled move is ignored. A deterministic withheld-rename test fails before this repair and passes after it. M724 restores the direct deletion and fails that test; M725 removes the stale-identity check and fails both pulled-move echo tests. Both controls apply and compile.

Continued typing also produced a deeper crossed-merge chain in a diagnostic full-suite run: one sequence survived only in a conflict copy. A successful merge using a virtual criss-cross base now defers when the local bytes differ from its authenticated recorded head. The normal push publishes the typing on that parent before another merge, so both devices merge published inputs. The regression requires no premature publication, intact typed bytes, then complete combined text, one server head and no conflict copies. Removing this deferral (M727) or reversing its digest comparison (M728) compiles and fails the witness. The existing three-level graph and memory bounds remain intact.

Two other asynchronous fixture barriers now wait for their actual completion: the nested-vault refusal before measuring notice quietness, and the new engine's restore probe before asserting its count. Their original assertions remain. Diagnostic traces and prior failures are retained outside the repo; they are not mutation kills or final native acceptance.

The follow-up's pinned make check passes 1,117 plugin tests, 144 core + 375 server + two CLI tests, 70 dashboard tests, 767 contracts and both secret scans, with 94.73% Rust line coverage. This includes the predecessor's new independent pull-persistence witness. Six separate pristine matrix lanes each pass the same 1,117 tests. All 673 mutation patches apply with zero fuzz; M152 retains its original crossed-base removal and M726 retains its save-only removal after re-cutting their context. M724, M725, M727 and M728 have one, two, one and three focused failing tests respectively, after successful builds. The current bundle is 530ca4aeb1fee893817f99110346226e7573714c542f3d303b45971e0f4b85aa. Complete mutation measurement and affected native acceptance remain pending.

The first final matrix stopped after 54 cases because M133 survived its complete 1,117-test run. Its ID-order-dependent witness is replaced as recorded in the review repair log. The strengthened 1.1.3 suite passes 1,120 tests in the pinned full gate; Rust, dashboard, contract, coverage and secret-scan results remain passing. Production source and the 530ca4ae bundle are unchanged. The partial matrix and its survivor are retained; the complete matrix must use the updated suite.