Test suites leave temporary directories behind on every run

[shibayan]

What happens

Several test suites create temporary directories under the OS temp directory and never remove them. They accumulate across runs on a developer's machine.

Measured on a Windows developer machine after ordinary use of the repository:

1560  afjs-fdry-*      (packages/foundry)
 288  afjs-replay-*
  61  afjs-bg-*        (packages/agentserver background tests)
  15  afjs-conv-*
  15  afjs-cont-*

A single vitest run packages/agentserver packages/foundry adds seven more (six afjs-fdry-*, one afjs-replay-*), so the count grows with every run rather than being bounded.

For contrast, the state-root isolation added for the response-store default (scripts/test-state-root.ts) removes its directory in afterAll and leaves nothing behind — afjs-state-root-* measures zero after a full run. The suites above predate it and have no equivalent cleanup.

Why this matters

• The directories hold whatever those tests wrote, including serialized transcripts and response envelopes, in the clear, indefinitely.
• On a long-lived machine the count reaches thousands, which slows directory listings and temp-cleaner passes.
• A test that reads "everything under the root" can observe another run's leftovers if a future change ever shares a root.

Required implementation

• Find every suite that creates a temporary directory and give it a cleanup that runs even when the suite fails.
• Prefer the pattern scripts/test-state-root.ts uses: name the directory rather than creating it eagerly, so a fully skipped test file leaves nothing (Vitest runs no afterAll for a file whose tests are all skipped), and remove it in afterAll with { recursive: true, force: true }.
• Where several tests in one file each make their own directory, one per-file root with per-test subdirectories is enough.
• Do not delete anything outside the directory the suite itself created.

Acceptance criteria

• A reproduction-first measurement records the leftover count before and after a full pnpm check, and the after-count for newly created directories is zero.
• Every suite that creates a temporary directory removes it, including when a test in that file fails.
• A fully skipped test file leaves no directory behind.
• The prefixes above (afjs-fdry, afjs-replay, afjs-bg, afjs-conv, afjs-cont) are each covered, and a sweep confirms no other prefix is left uncovered.
• Cleanup failures do not fail an otherwise passing suite.
• pnpm check passes.

Out of scope

Removing directories left by earlier runs. This issue stops the growth; existing leftovers are the developer's to clear.