> Outcome (added when this file was published, 2026-11-13): open, S4. Not fixed in 3.0.10:

From mnema-talk v3.3.13, rendered on this site.

ID: H-L2-06 Lane: 2 Severity: S4 (availability; needs a hostile or lossy relay to fork the room, or a relay that delays the committer's commits; heals, slowly) Status: claimed, reproduced by the reviewer in the simulator (phase 3 review of H-L2-05 / H-L9-02); not yet refuted

Claim: Two groups of a Talk that have the same member ids and the same epoch number cannot see each other, so a fork of that kind heals only by chance. A heartbeat names the group, the id, the epoch NUMBER and a time, signed with the member's leaf key (heartbeat.ts); it does not name the state of the tree or the epoch's key. A member whose own tree holds a leaf of the lowest id that signs with the same key as the other group's leaf (a package offered from inside a group keeps the signature key: SECURITY.md "A member that moved to the other group keeps its signature key") takes that id's heartbeats of the other group as signs of life "when the two epoch numbers happen to be equal" and starts nobody over (SECURITY.md, same paragraph). The only detector of another state, two heartbeats of the lowest id for an epoch that is not mine, needs the numbers to differ and the member's own state to stay unchanged between the two heartbeats, while both groups rotate and renew on their own schedules. In the simulator (rotation 40 s) the two groups end at the same number again and again and the room does not agree within the 240 s of the invariant.

Property: Invariant 1 of the simulator: once the relay behaves, the room holds one epoch and key within 240 s. SECURITY.md: "anything that has split is repaired by the existing start-over".

Repro: from msch128/fix-h-l2-05-l9-02 (9fe070ce3), in web/: SIM_CLOCKS=1 SIM_CLOCKS_VARIANT=steps SIM_CLOCKS_SEEDS=5 SIM_CLOCKS_OUT=/tmp/o5.jsonl npx vitest run src/lib/media/groupClocks.sweep.test.ts; cat /tmp/o5.jsonl The hostile row has "failed":true ("the room did not agree within 240000 ms: m7#4:14:... {1,7} | m1#8:14:... {1,7}"): two groups with the same members and the same epoch 14, different keys. The same end state on the HONEST relay: SIM_CLOCKS=1 SIM_CLOCKS_VARIANT=all SIM_CLOCKS_SEEDS=120 SIM_CLOCKS_OUT=/tmp/o120.jsonl npx vitest run src/lib/media/groupClocks.sweep.test.ts; cat /tmp/o120.jsonl ("m4#8:14 {2,4,6} | m6#9:14 {2,4,6} | m2#10:14 {2,4,6,11} | m11#11:14 {2,4,6,11}": two groups, epoch 14 in both, different keys.) Both seeds are in the failure budget (groupSweep.known.fixture.ts) with the comment "two groups of the same members that never merge" (seed 5) and "fails only with its forward step of 447 s" (seed 120). Measured: with the room given more time they DO heal: hostile 5 (variant steps) agrees 486 s after the relay stops, honest 120 (variant all) after 304 s, at the simulator's compressed lifetimes. At the lifetimes of production ratios (absence 800 s, heartbeat 200 s, rotation 1200 s, maintenance 20 s) the same seed and variant settle (a different history, because the timing changes the draws), so whether the production heal time is minutes or hours is not shown.

Minimal: (Read from the trace and from the code; the sequence is the reviewer's reading, not each step proven.) Seed 5, hostile, variant steps: m1 (the lowest id) reloads as m1#8 and joins a group through m3's welcome while the hostile relay holds its commits back from m7 for about 20 s. m7 meanwhile commits on its own branch (it takes m1 for the committer that reloaded: "the next one swaps the old leaf"), and when the delayed commits arrive at 75 s the two states are different groups of {1,7}. m7's welcome to m1#8 is ignored (m1#8's committer is itself: a welcome from an id that is not lower is no rival), and m1#8's welcome to m7 is ignored (the welcome's sender id equals the id of m7's committer, so it is not "from a lower id outside my group": rivalry is judged by id, not by leaf). From then on each group renews and rotates by itself (trace: commit@9[] ... commit@13[] from m7 and from m1#8 alternately, never taken by the other because neither verifies under the other's tree).

Reach: a relay that delays or drops the committer's commits for a member for longer than the hung-committer skip while the lowest id reloads, or a lossy honest relay (seed 120 is an honest relay) with several reloads. No attacker needed.

Effect: two parts of a Talk with different keys: no media between them (each part hears nothing of the other) until a membership change, a rotation that makes the numbers differ for long enough, or the zombie leaf's removal. The simulator shows 5 to 8 minutes of compressed time.

Not shown: the production heal time; the exact origin of the fork of seed 120 (the sleep check of a member that left 20 s later changes the history, but the end state is the same kind as seed 5's; with the wall step made a no-op seed 120 passes, with the step applied but the sleep mark removed it still fails, so it is the member's own wall clock read by the sleep check that moves the history, not the mark the simulator sets); whether a real browser forks this way.

Fix: direction only: let a heartbeat (or the capability message each member sends with it) carry a short tag of the epoch (for example the first 8 bytes of a hash of the epoch authenticator): two states with the same number then differ in the tag, the member that sees the lowest id's tag differ starts over as for any other state. A wire change: old clients ignore the extra bytes if the version byte allows; a capability bit would gate it. Alternatively compare by leaf instead of id when deciding whether a welcome is a rival one (the welcome built on the package I offered last, from an id equal to my committer's, with a leaf of mine in it). Protocol code under web/src/lib/media: protected area, owner decides.

Reviews: reviewed once by the reviewer of fix-h-l2-05-l9-02 (see doc/security/reviews/2026-11-hunt/h-l2-05-h-l9-02.md); reproducer and refuter: open.

Something wrong on this page? Tell us.