fix(guard): scan a command substitution in its own quote context (169)

[karlkfi]

Fixes #169.

What

strip_heredoc_bodies tracked quote state flat, so a heredoc opened inside a
$(…) or backtick substitution was invisible whenever the substitution sat in
double quotes. Bash opens a fresh quoting context there; the hook did not, so it
stayed in in_double from the opening " onward and the << branch was never
reached.

The body then survived into shlex. With an odd number of " in it the parse
raised No closing quotation, _analyze_command returned its defer tuple, and
nothing in the command was analyzed — including a guarded outside-workspace
path sitting in plain sight at the end of the line:

git commit -aqF "$(cat <<'MSG'
he said "hello
MSG
)" && cp secrets.txt /etc/newfile

That is the shape of an ordinary multi-paragraph commit message. Measured on the
tip before this change, a quote character in an unrelated commit message turned
a hard deny into silence:

trailing command	no heredoc	even-quote body	odd-quote body
cp secrets.txt /tmp/q169-fake	deny	deny	silent
tee /tmp/q169-fake	deny	deny	silent
rm -f /var/tmp/q169-other/x	deny	deny	silent
mv notes.md ~/q169-fake.md	ask	ask	silent

How

Push the enclosing quote state on $( and on a backtick, scan the body
unquoted, restore it on the terminator. Paren depth is tracked per context so a
subshell's ) does not end the substitution early.

Recursing into $(…) alone is not sufficient, as the issue notes:
_scan_dollar_paren tracks quotes flat too, so the body's " swallows its
closing paren and there is nothing to recurse into. Stripping the body first
hands that scan a balanced string, so it needs no change — the fix stays in one
function.

Backticks are the same bug one character away ("cat <<'M' … M" was silent
too), so both openers are handled rather than only the reported one.

Unchanged by design: $((…)) still reads as arithmetic and never arms a
delimiter; a << in quotes or in a # comment is still not a heredoc; and the
expanded list still collects unquoted-delimiter bodies wherever they sit, so
the Q35 command-substitution scan is unaffected.

Testing

python3 scripts/run-tests.py — 1311 tests, OK (8 skipped).

12 new tests. Nine assert the moved behavior and were confirmed to fail against
the pre-fix script, each with the reported symptom (expected a decision, got defer); three are regression fences that hold on both sides, pinning the cases
where the fresh context must not reach — "$((1<<2))", "$(grep -c '<<' f)",
"$(echo "a<<b")".

Also checked by hand that the change only ever removes heredoc bodies: for a
dozen substitution-heavy commands with no heredoc in them, strip_heredoc_bodies
returns its input unchanged.

Docs

README.md — the "How it works" tokenization step and the heredoc entry under
Limitations both now state that a substitution body is scanned in its own
quoting context.

Known residual

A case pattern's ) inside a quoted substitution still pops the context
early, so this narrower shape stays silent (measured on this branch):

echo "$(case $x in a) cat <<'M'
he said "hi
M
;; esac)" && cat /etc/q169-fake

Closing it needs case/esac keyword tracking, and mis-detecting those
over-strips — the unsafe direction — so it is not a follow-on one-liner.
Queued as Q81 rather than bundled here.

Release note

A heredoc inside a quoted command substitution — git commit -F "$(cat <<'MSG' … MSG)" — no longer hides the rest of the command: an odd number of quotes in the message body used to abort the parse, silently skipping an outside-workspace path later on the line.