feat(hosts): add a cron bridging helper for macOS/Linux, wire claude_code - #595

The two tests locking the docs' pipeline prompt to
bridging_pipeline_prompt() grepped for the inline `-p 'Run the memU`
cron line, which the previous commit replaced with a prompt-file block.
Anchor on the prompt line itself (the single-line fence written to
pipeline-prompt.txt) instead — same verbatim-equality guarantee, new
location. Also refresh prompt.py's module docstring, which still
described the Unix mechanism as pasting the prompt into the crontab.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ridge script

cursor-agent refuses headless runs in an untrusted directory (Workspace
Trust Required, exit 1) — field-verified in memU#571, which flagged the
Unix cron path as still missing the flag. The wrapper now cds into
~/.memu/hosts/cursor and invokes cursor-agent --trust, so trust lands on
memU's own tree rather than wherever cron starts, mirroring the wired
Windows template. Never --yolo.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…try length, document lock tradeoff, make hermes native-first

Four review points, all pre-merge cleanups:

- Filenames converge with the shipped Windows helper: bridge-prompt.txt
  (was pipeline-prompt.txt) and bridge.log (was bridging.log), so the
  planned Unix schedule helper and every cross-platform doc carry one
  naming scheme instead of two forever. bridge.sh keeps its name — the
  extension already distinguishes it from memu-bridge.ps1.

- New regression gate: a parameterized test over the four Unix guides
  asserts every cron entry line stays under 512 chars. The prompt-lock
  tests catch content drift but not re-inlining — this catches the
  actual bug class (memU#591) at test time instead of on the next Mac.
  Generic's entry moved into a fenced block so the gate sees it.

- The 3h lock-reclaim tradeoff is now spelled out in the script comment:
  a legitimate >3h run loses its lock to the next tick and can double-run,
  accepted deliberately over a crashed run wedging the schedule forever.

- Hermes: native cronjob is now the recommended path in so many words;
  raw crontab is explicitly fallback-only, with the history note that no
  install is known to have ever succeeded through it (the nonexistent -p
  flag survived in the guide precisely because nobody ran that path).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…code

The Unix half of the schedule helper (NevaMind-AI#546 shipped the Windows half):
`memu-claude-code schedule {install,verify,status,uninstall}` now works on
macOS/Linux against cron. install writes bridge-prompt.txt + a locked
bridge.sh (PATH baked in, cd into the host tree) and registers one short
crontab entry — replacing, never duplicating, any previous entry for the
host, including a legacy inlined one (memU#591), which install thereby
migrates. verify proves the entry can run without waiting for a tick;
uninstall removes both layouts and keeps the run log.

The scheduling package fronts both backends behind one verb surface,
picked by platform. The auth/resolution gates are imported from
windows.py — they were always platform-neutral. claude_code's guide now
leads with the helper and keeps the manual registration as the fallback.

Real-machine check: `schedule status` against a live install correctly
reports the crontab entry and a held lock mid-run. 237 tests, mypy and
ruff clean.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>