Summary

Follow-up to #626. The Ollama branch of _step_embedding was migrated to the new fail_step() mechanism in #661, but two other interactive steps still perform filesystem side effects without any try/except. They don't silently advance (the original #626 bug pattern) — they crash with an uncaught stack trace, which is the symmetric failure mode and equally surprising for the user.

Sites

1. _step_memory_dir (packages/memtomem/src/memtomem/cli/init_cmd.py:638)

if not memory_path.exists():
    create = nav_confirm(f"  '{memory_dir}' doesn't exist. Create it?", default=True)
    if create:
        memory_path.mkdir(parents=True, exist_ok=True)   # <-- uncaught
        click.secho(f"  Created {memory_path}", fg="green")

If mkdir fails (PermissionError, OSError, ENOSPC), the wizard exits with a Python stack trace. The user has to read the trace, fix the underlying issue, then re-run mm init from the top.

Suggested fix: wrap the mkdir in a while True: try / except (PermissionError, OSError) as exc: fail_step(f"Could not create '{memory_path}': {exc}", retryable=True) block so the user can fix permissions / pick a different path / quit.

2. _step_settings (packages/memtomem/src/memtomem/cli/init_cmd.py:889-894)

canonical = project_root / CANONICAL_SETTINGS_FILE
canonical.parent.mkdir(parents=True, exist_ok=True)   # <-- uncaught
if not canonical.exists():
    canonical.write_text(                              # <-- uncaught
        json.dumps({"hooks": {}}, indent=2) + "\n",
        encoding="utf-8",
    )

Same pattern: mkdir and write_text can both raise. Crashes on read-only project root, full disk, etc.

Suggested fix: same shape as above — try/except + fail_step with retryable=True.

Out of scope (intentionally)

The "warn-and-save" patterns in _step_embedding (ONNX/Ollama/OpenAI client missing, kiwipiepy missing in _step_language) are documented UX — the wizard saves the config so the user can install the extra later. Those are not bugs.

Why a separate PR

Per the project's "one focused PR per change" convention, the Ollama branch fix in #661 deliberately did not touch other steps. Once this PR lands, the fail_step adoption is consistent across every step that performs a side effect.

Acceptance

• mkdir / write_text failures in _step_memory_dir and _step_settings surface a fail_step prompt instead of a stack trace.
• Tests pin: simulate mkdir raising PermissionError → assert the prompt appears, r retries the action, b returns to the previous step, q cancels via WizardCancel (exit 0).