[pchiusano]

Overview

Fixes a flaky Windows failure of the MCP runner test "a branch created by an earlier command can be named". It removed #6324 from the merge queue (run):

EXCEPTION!!!: ...\ucm-mcp-runner-test-31cceb5edf4b0451\.unison\v2\unison.sqlite3-wal:
removeDirectoryRecursive:...:DeleteFile "...\unison.sqlite3-wal": does not exist

The cause is in ucm, not the test:

• Switching branches (Cli/Monad.hs) calls preloadProjectBranch, which called preloadBranch.
• preloadBranch forked a thread with forkIO that nobody tracked. The thread opens its own SQLite connection to load the branch into the cache.
• The test's transcript switches to a branch it just created. If the transcript ends while that thread is still loading, the codebase closes with the connection still open.
• The test then deletes the codebase directory while that last connection closes. Closing the last connection makes SQLite delete the WAL file, so the file vanishes in the middle of the delete. On Windows that is an error.

Implementation approach and notes

• sqliteCodebase opens a ki scope for the codebase's lifetime (withPreloads). The scope is inside the codebase lock.
• preloadBranch forks into that scope. When the codebase closes, ki cancels any preload still running and waits for it to finish. No connection outlives the codebase.
• A preload catches its own synchronous exceptions. A failure only means the branch isn't cached, as before, and doesn't propagate to the thread that opened the codebase.
• parser-typechecker now depends on ki, which unison-cli already uses.

Test coverage

• The existing test is the regression test. The race needs Windows to fail, so I couldn't reproduce it on macOS. The evidence for the fix is the explanation above.
• The MCP runner tests pass on macOS, and so do the tests, transcripts, formatting and migration proofs.

Loose ends

None.

[pchiusano]

Cc @mitchellwrosen

Ki looks great btw I haven't really used it before.