Summary

Closes #421.

_step_provider_dirs_auto sits between _step_memory_dir and _step_mcp
but only emits a detection banner — it has no user prompt. Pre-fix, "b" at
_step_mcp decremented onto this silent step, which re-emitted its banner
and advanced straight back to _step_mcp, so "b" appeared to do nothing.
Same "boundary can't be crossed" shape as #371 / #418 but one step further
along the flow.

Fix

• wizard.py: new silent_step decorator sets a _silent_in_back_nav
  attribute on the step function. run_steps's StepBack handler now walks
  back past any consecutive silent steps until it reaches an interactive one
  (or echoes (already at first step) if none precedes the current
  position). Forward direction is unchanged — silent steps still run their
  banner on the way through.
• init_cmd.py: decorate _step_provider_dirs_auto with @silent_step
  and document the skip contract in the docstring. It is the only silent
  step today; the decorator is exposed generically from wizard.py so
  future silent steps opt in the same way.

Acceptance criteria (from #421)

• b at _step_mcp lands the user back at the previous prompt
  (memory-dir), not on a silent step.
• Silent step side effects (banners) don't re-fire on a back-pass-through.
• New test that sequences CliRunner().invoke(init, input="...") and
  asserts the memory-dir prompt re-appears in the output.

Test plan

• tests/test_wizard.py (new) — 9 unit tests covering the skip mechanism
  directly: single silent skip, multi-silent skip, forward-only flow,
  back-at-index-0, silent-at-position-zero, state persistence across back-nav.
• TestBackNavThroughSilentStep in test_init_cmd.py — CliRunner regression
  that pins the real wiring. Memory Directory header must appear twice in
  captured output (initial visit + post-back re-entry) and the banner text
  must be absent from the slice between the first mcp prompt and the second
  memory-dir header.
• pytest packages/memtomem/tests -m "not ollama" → 2233 passed, 46
  deselected.
• ruff check + ruff format --check + mypy init_cmd.py wizard.py → clean.

Out of scope (tracked separately)

• Step-number display drift (init: step_header(N, ...) hardcodes step number — preset flows show misleading positions #420) — step_header still hardcodes its
  number regardless of the flow. Orthogonal UX polish; next PR.
• --preset <name> false-advertises (b: back) (init --preset <name>: memory-dir step advertises (b: back) but b has nothing to return to #422) — memory-dir is
  the only step there, so "b" is legitimately a no-op but the hint lies.
  Cosmetic twin, good-first-issue.