Timeline

Phase 2 only ever proved convergence after the network was fully healed, so a protocol that converges on a perfect wire but livelocks under sustained loss would have passed. A new degraded phase holds a fixed, never-healed drop and duplication rate on both links and drives a fresh session through the same safety, convergence and termination oracles, within a window scaled to the loss rate: a round makes progress only if both its request and reply arrive, so the expected round count grows as one over the delivered-fraction squared, times a safety factor, with the derivation in a recomputable comment. It converges with wide margin -- across many seeds the worst run finished in a small fraction of the window at every rate tested -- so there is no livelock. Pinning the window below the real need trips the oracle, so the check is not vacuous. The gate runs one modest rate; the soak sweeps a range. check-in: b8ef4a68e5 user: dmestas tags: trunk

The loss-scaled convergence window used divCeil with a catch unreachable for the divide-by-zero arm, which the precommit guard flags outside tests. The delivered-percent denominator is provably at least one, so plain integer ceil division gives the same bound with no error union. closed check-in: 0fbc9225bf user: dmestas tags: h7-degraded

The sync sweep proved convergence only after the network fully healed, so a protocol that converges on a perfect wire but livelocks under sustained loss would still pass.

Phase 2b runs a fresh session over a wire held at a fixed drop/dup rate, never healed, asserting the same convergence and termination oracles within a window scaled from the loss rate: expected rounds are the clean window over (1 - p_loss) squared per direction, times a safety factor, recomputed in a comment rather than a magic number. No corruption is injected -- a corrupted body is already a lost round.

The protocol converges comfortably: over 200 seeds the worst run took 7, 13, and 18 rounds at 10, 25, and 40 percent loss, against windows of 160, 228, and 356 -- no livelock. The gate joins at 10 percent; the soak sweeps the wider rate set. Pinning the window below the real need trips the liveness oracle, so the check is not vacuous. The liveness-under-faults discipline follows TigerBeetle's VOPR. check-in: 464b82c888 user: dmestas tags: h7-degraded

The working tree -- the one place a version-control tool can lose a user's data -- had no crash simulation; only the log store did. A new faulty directory layer tears and freezes writes at a seeded crash point, and a sweep replays an update across every crash cut, reopens, and checks four invariants. Three hold: an untracked file the update must preserve is never touched, a crashed update recovers to its target through the undo journal, and the checkout record never names a check-in whose tree did not finish materializing. The fourth does not, and that is the point of the harness: the payload write lands in place rather than through the write-temp-and-rename the metadata already uses, so a crash mid-write can leave a managed file spliced from old and new bytes -- torn, but recoverable. The gate runs the three that hold; the fourth is pinned by a test asserting the splice IS produced, so it flips the day the write is made atomic. Crash-point enumeration follows TigerBeetle's storage model. check-in: 5401ce245f user: dmestas tags: trunk