Nightly stress failure: multi-process

[github-actions]

• Workflow: Nightly stress
• Run: https://github.com/dolthub/doltlite/actions/runs/32937168901
• Commit: 10170ed82c1b12414db8d1b29d2fe9ea2a72fd88
• Ref: master

Nightly stress soak failed in multi-process.

Full log is attached as the nightly-stress-multi-process artifact on the run.

Reproduce locally after a checked build:

cd build
DOLTLITE_MP_COMMIT_OPS=2000 ./multi_process_test
./multi_process_gc_test
./multi_process_merge_rebase_test

Last 80 lines of the soak log:

=== Multi-Process Merge & Rebase Concurrency Tests ===

--- Test 1: 8 processes merging distinct branches into main ---
--- Test 2: merge into main racing concurrent main commits ---
--- Test 3: rebase racing a dolt_gc() loop ---
--- Test 4: rebase racing a checkout loop ---
--- Test 5: concurrent --continue vs --abort ---
  continue wins=6 abort wins=6 bad=0
--- Test 6: concurrent --continue after default reopen ---

=== Results: 46 passed, 0 failed out of 46 tests ===
=== Multi-Process Merge & Rebase Concurrency Tests ===

--- Test 1: 8 processes merging distinct branches into main ---
--- Test 2: merge into main racing concurrent main commits ---
--- Test 3: rebase racing a dolt_gc() loop ---
--- Test 4: rebase racing a checkout loop ---
--- Test 5: concurrent --continue vs --abort ---
  continue wins=4 abort wins=8 bad=0
--- Test 6: concurrent --continue after default reopen ---

=== Results: 46 passed, 0 failed out of 46 tests ===
=== Multi-Process Merge & Rebase Concurrency Tests ===

--- Test 1: 8 processes merging distinct branches into main ---
--- Test 2: merge into main racing concurrent main commits ---
--- Test 3: rebase racing a dolt_gc() loop ---
--- Test 4: rebase racing a checkout loop ---
--- Test 5: concurrent --continue vs --abort ---
  continue wins=5 abort wins=7 bad=0
--- Test 6: concurrent --continue after default reopen ---

=== Results: 46 passed, 0 failed out of 46 tests ===
=== Multi-Process Merge & Rebase Concurrency Tests ===

--- Test 1: 8 processes merging distinct branches into main ---
--- Test 2: merge into main racing concurrent main commits ---
--- Test 3: rebase racing a dolt_gc() loop ---
--- Test 4: rebase racing a checkout loop ---
--- Test 5: concurrent --continue vs --abort ---
  continue wins=6 abort wins=6 bad=0
--- Test 6: concurrent --continue after default reopen ---

=== Results: 46 passed, 0 failed out of 46 tests ===
=== Multi-Process Merge & Rebase Concurrency Tests ===

--- Test 1: 8 processes merging distinct branches into main ---
--- Test 2: merge into main racing concurrent main commits ---
--- Test 3: rebase racing a dolt_gc() loop ---
--- Test 4: rebase racing a checkout loop ---
--- Test 5: concurrent --continue vs --abort ---
  continue wins=5 abort wins=7 bad=0
--- Test 6: concurrent --continue after default reopen ---

=== Results: 46 passed, 0 failed out of 46 tests ===
multi_process_merge_rebase_test run 210
=== Multi-Process Merge & Rebase Concurrency Tests ===

--- Test 1: 8 processes merging distinct branches into main ---
--- Test 2: merge into main racing concurrent main commits ---
--- Test 3: rebase racing a dolt_gc() loop ---
--- Test 4: rebase racing a checkout loop ---
--- Test 5: concurrent --continue vs --abort ---
  continue wins=6 abort wins=6 bad=0
--- Test 6: concurrent --continue after default reopen ---

=== Results: 46 passed, 0 failed out of 46 tests ===
=== Multi-Process Merge & Rebase Concurrency Tests ===

--- Test 1: 8 processes merging distinct branches into main ---
--- Test 2: merge into main racing concurrent main commits ---
FAIL: merge_race_feat_rows_present
--- Test 3: rebase racing a dolt_gc() loop ---
--- Test 4: rebase racing a checkout loop ---
--- Test 5: concurrent --continue vs --abort ---
  continue wins=4 abort wins=8 bad=0
--- Test 6: concurrent --continue after default reopen ---

=== Results: 45 passed, 1 failed out of 46 tests ===
completed 211 multi_process_merge_rebase_test runs

[timsehn]

Fixed by #2436, merged.

Root cause, captured live with instrumented soaks: a fast-forward merge racing a concurrent commit could publish a poisoned working set. The head confirm inside CompareAndAdvanceBranch reopens the store when a peer changed it, and that reload replaced the merger's live session catalog with the writer's freshly written working set; PersistWorkingSetWithHash(NULL) then bound that stale content to the newly advanced head, the staleness gate accepted it, and the writer's next commit made the loss permanent (the preserved store shows feat work in main's ancestry with zero feat rows).

Fix: advancing with no explicit working catalog persists the advance's own catalog rather than the live session. Verified with three CI soak runs on the fixed build — 7,200 nightly-warmup loops, zero failures, where the unfixed rate (~1/300 per loop, reproduced twice) predicts ~24. The temporary soak rig branch has been deleted; the nightly stress soak remains the sentinel.