Shell installer tests has been failing on main, and failing in the worst way:
the file printed its header and exited 1 with no FAIL line, because two
separate things broke and both of them are silent.

#8805 added the uv download markers, which put the wrapped command inside
`{ ...; } | _uv_download_markers ...`. The verbose path was already a pipe;
the quiet path, which is the one this file exercises, was not. Two
consequences:

1. This file hand-copies five functions out of install.sh and stubs five
   more. It never learned about _uv_download_markers, so the last stage of
   that pipeline was "command not found", every wrapped command looked like
   it exited 127, and `set -e` ended the run. Extract the real function
   rather than stub it, so the test still exercises the pipe the installer
   actually runs.

2. The command now runs in a subshell, so _test_command's `_test_attempt`
   increments were discarded. It failed on every attempt instead of
   succeeding on the second, which is precisely the retry-recovers path the
   first assertion is about. Move the counter into a file.

Both are needed; each alone still fails, at a different assertion.

#8805 is right on both counts. Streaming output through a marker filter
requires a pipe, and real installs run external commands, so a subshell costs
them nothing. It was the test's harness that had gone stale.

Also add `|| true` to the two `grep -c` calls that gate the first
assertion. grep -c exits 1 on a zero count, which under `set -e` aborts with
no output. That is why this read as silence rather than as a failure, and the
same guard is already applied to the grep -c calls later in the file.

tests/sh: 47 files, all passing.

The second trap replaces the first, so naming the attempt file only in the
first one leaked it: one stray temp file per run, measured.

My earlier commit message claimed the grep -c guard was already applied to the
rest of the file. That was wrong: lines 231, 326 and 368 were still unguarded,
so a zero count there aborts under set -e before the FAIL line it gates.

Worse, the guard I did add only covers the case where a wrapped command
succeeds and emits the wrong markers. The dominant path is the command itself
returning non-zero: under set -e that aborts at the call, before any assertion
runs. A broken installer retry therefore still reproduced the exact
header-then-silence symptom this file was fixed for.

Guard the three grep -c sites, and give the four commands that must succeed an
explicit failure message instead of letting set -e end the run mutely.

Each path mutation checked, silence to a named failure:

  retry never recovers       -> FAIL: recovered retry returned non-zero
  rollback_substep renamed   -> FAIL: successful Unix rollback output ...
  UNSLOTH_TAURI_MODE renamed -> FAIL: Unix installer does not pass Tauri mode

tests/sh: 47 files, all passing, no temp files leaked.