ENG-1646: self-heal scratchpad venv creation instead of retrying an identical failure

[torrmal]

Summary

Fixes ENG-1646.

uv venv normally symlinks a new venv's bin/python straight to the base interpreter. On at least one real machine, when invoked from inside the packaged Cowork desktop app's process tree, uv instead wrote bin/python as a freshly-copied, merely ad-hoc-signed launcher — missing its own libpythonX.Y.dylib — and macOS's AMFI killed it outright the moment anything tried to run it:

kernel: (AppleMobileFileIntegrity) AMFI: '.../bin/python' has no CMS blob?
kernel: (AppleMobileFileIntegrity) AMFI: '.../bin/python': Unrecoverable CT signature issue, bailing out.

Because the failure is deterministic per environment (not transient), all 3 of _ensure_venv()'s retries repeated the identical uv venv call and failed identically — so restarting the app, renaming the scratchpad, or waiting and retrying never helped. The full repro/diagnosis is in the Linear issue.

Changes

• _create_venv() now checks whether uv actually produced a symlink for bin/python; if it instead wrote a copy, _repair_copied_launcher() replaces it with a direct symlink to the resolved base interpreter. Every symlinked interpreter observed in testing worked; every copied one was killed by AMFI.
• _ensure_venv()'s retry loop now escalates strategy instead of repeating the same doomed call: the final attempt bypasses uv entirely and falls back to stdlib venv.create(symlinks=True), which has been reliable in every environment this was tested against.
• The venv directory is no longer nuked after the final failed attempt — it's left in place for inspection instead of erasing the only evidence of what went wrong.
• Failures now carry real diagnostic detail (_diagnose_broken_interpreter(): symlink vs. copy, file size, and codesign status on macOS) instead of going silent whenever _last_verify_error was never set (e.g. the interpreter file simply doesn't exist).

Test plan

• uv run pytest tests/test_local_venv_provisioning.py -v — 23 passed (6 new, covering the repair, the escalation order, directory preservation, and the new diagnostic detail)
• uv run pytest tests/test_scratchpad.py tests/test_chat_scratchpad.py tests/test_scratchpad_flavor.py tests/test_scratchpad_utf8.py tests/test_scratchpad_watchdog_contracts.py tests/test_scratchpad_observer_dispatch.py -q — 211 passed, no regressions
• Manual verification on the machine where this was originally observed failing in production

🤖 Generated with Claude Code