ci/spec22-gems: fix the env -i continuation split that dropped TEBAKO_RUNTIME_IMAGE (incident 13 round 5)

[ronaldtse]

Incident 13 round 5 — root cause of the run-32188296332 all-legs failure

The instrumented rerun failed all four jailed legs with a NEW signature:

`RubyGems' were not loaded.
`error_highlight' / `did_you_mean' / `syntax_suggest' were not loaded.
A:/probe/probe.rb:66:in '<main>': undefined method 'clear_paths' for module Gem (NoMethodError)

plus a misleading driver debug line mounted what=failed to mount image ... at=A:/
(that line is the SUCCESS path — what is the error-context string; the payload
mount worked, as the [tfs] open: A:/probe/... lines prove).

Root cause

#89 (e318712) placed a 5-line comment block inside the backslash-continued
env -i list in run_probe. A comment line inside a continuation terminates the
command at that point (the \ splices the comment line in; the comment ends at its
own newline), so run_probe executed two commands:

1. env -i MSYS2_... TEBAKO_RUNTIME_IMAGE=... — no command argument: prints the
   env and exits 0, invisible under set +e.
2. TEBAKO_JAIL=... TEBAKO_DEBUG_TFS=1 "$RUNTIME_EXE" --tebako-image ... — running
   with the ambient msys environment and without TEBAKO_RUNTIME_IMAGE,
   TEBAKO_HOME, HOME, APPDATA, USERPROFILE, TMP*, the System* baseline,
   and the env scrub.

With TEBAKO_RUNTIME_IMAGE unset the driver early-returns and the env image never
mounts, so rubygems and the default gems cannot load — exactly the prelude warnings
above — and gem_prelude's partial Gem module dies at probe.rb:66. The
TEBAKO_HOME loss also explains the r9 artifact's missing journal.log (the store
journal went to the runner's real ~/.tebako; the upload path was correct).

Verified locally with a same-shape repro: comment-in-list splits the command
(ambient env leaks, listed vars lost); comment-outside delivers every assignment
under the scrub. bash -n clean. The one other comment-after-continuation site in
ci/ (run-msys.sh:94) is inside a comment block — harmless.

tebako#417 exonerated

Its traces are pure diagnostics; the r9 logs show the closure-walk verdicts working
(styles-tree materialize cache hits). The traces + this flag's passthrough are
exactly what made the mis-parse diagnosable in one read.

The tripwire (second commit)

probe.rb now prints the three handoff var presences (TEBAKO_RUNTIME_IMAGE,
TEBAKO_HOME, HOME) immediately after the PROBE gem-loaded canary — pure ENV
reads, no rubygems dependency, never gates. A lost handoff names itself in the
first lines of every leg's proof log instead of a full CI cycle later.

Expected effect on the factory dogfood

Restores the r8 environment AND keeps the TEBAKO_DEBUG_TFS traces — the next
factory run should show the r8 leg outcomes plus the closure-walk [tfs] verdicts
for the original sassc 126, which round 4 could not see.