mirror of
https://github.com/PerpetualSoftware/pad.git
synced 2026-10-26 22:37:03 +00:00
xarmian
5caf11f829
* fix(collab): a REST content replacement fences tabs seeded from the old body, as a restore does (BUG-3746) A prune-and-replace content write (content route and patch route, via PruneAndApply) pruned the op-log and replaced items.content, but stamped no seed boundary. A browser that fetched body A before the write and joined after it, with since=0 and its old content_seq, was admitted, seeded A into the empty document and flushed it over B. Join already force-refreshes a client whose content_seq predates the last restore (BUG-2264); replacements now stamp that same boundary: MarkRestoreBoundary on the write (durable last_restore_seq, in its tx) and SetLastRestoreSeq in memory, under the item lock Join takes, so no join lands in between. Claude-Session: https://claude.ai/code/session_01Rj4CMyK2EqxerMihbGs3Kr * test(collab): the BUG-3746 boundary test uses a collab-enabled server (BUG-3746) testServer has no collab manager, so the PATCH took the ordinary write path, never reached PruneAndApply, and the test failed with the fix in place (CI on #2134). Production always runs with collab; the test now does too. The seeded op-log row is dropped: the boundary does not depend on it. Claude-Session: https://claude.ai/code/session_01Rj4CMyK2EqxerMihbGs3Kr * test(collab): the patch route has its own seed-boundary leg (BUG-3746) The patch route prunes and replaces through its own PruneAndApply call, wired separately from the content route, so a future edit dropping fenceSeedsBefore there would pass the content-route test. Its twin asserts the boundary moves past the seeded body's seq to the patch's. Claude-Session: https://claude.ai/code/session_01Rj4CMyK2EqxerMihbGs3Kr * test(collab): a REST replacement admits a tab seeded at the new seq and refuses one seeded before it, through a real join (BUG-3746) Heron, GO-blocking on #2134: the fence must not over-fence, or every fresh tab after a REST replace would be force-refreshed, possibly in a loop. Per route (CONVE-19), after the replacement a tab joining with content_seq = the new seq gets op_log_cursor, and one at seq - 1 gets force_refresh. Mutant: the in-memory boundary stamped one past the write fails both admit legs ("the fence over-fences"). Counterfactual: with the fence removed, both refusal legs fail. Claude-Session: https://claude.ai/code/session_01Rj4CMyK2EqxerMihbGs3Kr