#664 What Muse's read of #634 left: the mutant lists read in order, a comparison's own operator, the head commit's trailer, the merge base (#589)

Part of #589. This is the follow-on to #634, the advisory check of a changed line of Python no test notices. #634 merged before Muse could read it: Muse was out of quota until 2026-11-09. Muse Wonder Full read it after the merge, at 60316bb8, and found four majors. All four are real, and each is fixed here.

Decision 0166 gains Boundaries rows only; its words are not changed. No protected file is touched, and no record is added.

Muse's four findings on #634, fixed

Finding	The fix	Held by
Any call in the mutant lists was data, so an entry that read the world, such as os.getenv('X', 'old'), escaped the check	An entry is data only if every name it reads was bound before the list to a text, a literal value or a helper the module defines; see below	test_a_call_in_the_mutant_lists_is_data_only_of_the_modules_own_helpers, 27 mutants
A comparison's fallback mutant took the first operator the line showed, even one inside a string, so a test that read the text could kill a mutant that changed nothing	The operator is found between the comparison's two sides. A comparison it cannot swap, or one on another line of the statement, is passed over	test_a_mutant_is_made_of_an_added_line_and_never_of_a_quiet_one, 5 mutants (<, >, >=, the guards)
Nothing held that the trailer is read from the head commit alone	A trailer in an earlier commit of the change is never read, and the commit asked for is the one read, not the checkout's HEAD	test_a_trailer_is_read_from_the_head_commit_alone, 2 mutants
Nothing held that the change is read from the merge base	Main moving on after the branch left it adds no hunk to the change	test_the_change_is_read_from_the_merge_base_and_not_from_main_as_it_moved_on, 1 mutant

What is data in the mutant lists, now

The check skips a hunk only when it is data: an entry of ISSUE or MUTANTS that decides nothing. The fix rounds found many routes by which an entry that reads the world passed for data. The rule is now stated in terms of what is allowed, not what is forbidden:

• A name is data only if a statement before the list bound it to one of these:
  • a text;
  • a value that is itself data;
  • a function the module defines, with no decorator.
• A name bound any other way is code:
  • bound twice;
  • bound by an import, wherever the import stands;
  • bound by a walrus;
  • named by a global;
  • changed by a method call, a store into it or a del;
  • left unbound.
• A call of globals, locals, vars, exec, eval, setattr or delattr makes every name code.
• Calls:
  • a call is data only of such a helper, or of replace on a text;
  • a helper's return counts as a text only if the helper says -> str;
  • a dunder attribute is code.
• The real lists read exactly as before: data_lines of tools/mutants.py is the same under main's untested.py and this one.

Residual, stated, not fixed: a contrived evasion is left out. One example is a nested scope whose own name masks the module's. The check is advisory, and the record already says the head's own tests can forge its counts. The hunk that reads the world sits outside the lists and is still checked.

A deliberate narrowing: a text method other than replace, such as .strip() or .format(...), is now code where main took it for data. No entry of the real lists changes.

Checks

• make verify on an export of the head: 1716 tests OK, at 2b47efa.
• Every mutant of tools/untested.py, run by tools.mutants on an export of the head: 83 of 83 killed at 2b47efa. 35 of them are new.
• All mutant targets occur once in their files, and no description repeats.

Reads

Argon, Muse Wonder Full, the Claude review and the security pass read the first head, eb184e4. The Claude review read each fix round again: fe00bf1, c0065ce and 2424df1. Argon read 2424df1. The last two rounds, 5626728/dd25df3 and 2b47efa, close what those reads found, and add a case for a mutant that survived; no reader read them again.

Reviewer	Head	Unread parts	Blocking	Major	Minor	Info
agy	eb184e4	none	0	2	0	0
muse	eb184e4	none	0	2	3	0
claude-review	eb184e4	none	0	0	3	0
claude-security	eb184e4	none	0	0	1	0
claude-review-r2	fe00bf1	none	0	2	2	0
claude-review-r3	c0065ce	none	0	1	3	0
agy-r4	2424df1	none	0	8	0	0
claude-review-r4	2424df1	none	0	0	1	4

Finding	Severity	Criterion	Where	What	Disposition
agy#1	major	PR-3	tools/untested.py:200	The receiver guard _literal(func.value, helpers) on .replace calls has no negative test and no mutant	Fixed in fae6b69: replace is data only on a text; ROOT.replace is code.
agy#2	major	PR-3	tools/untested.py:200	_literal treats os.replace(...) and any module's replace function as literal data	Fixed in fae6b69: os is a name an import binds, so os.replace is code.
muse#1	major	PR-3	tools/untested.py:162	Helper detection ignores binding order, so a shadowed name calling os.getenv is data	Fixed in fe00bf1: the module is read in order; a name bound twice is code, and a list sees only what was bound before it.
muse#2	major	PR-3	tools/untested.py:200	Any object's replace, including os.replace, is treated as mutant-list data	Fixed in fae6b69, as agy#2.
muse#3	minor	PR-3	tools/untested.py:200	No test holds the replace receiver check against a deciding receiver	Fixed in fae6b69 and fe00bf1: open('x').read().replace and os.replace are held as code.
muse#4	minor	PR-3	tools/untested.py:484	No test holds the > and >= entries of the comparison SWAP table	Fixed in fe00bf1: > and >= are swapped in the test, with a mutant each.
muse#5	minor	PR-3	tools/untested.py:521	Trailer test never calls trailers() for a commit other than the current HEAD	Fixed in fe00bf1: the earlier commit is asked for after the fix round too, with a mutant.
claude-review#1	minor	PR-3	tools/untested.py:191	An entry of the mutant lists that reads the world through a subscript or attribute is still data	Fixed in fae6b69: a comparison it cannot swap is passed over, with a case and a mutant.
claude-review#2	minor	PR-3	tools/untested.py:498	No test holds the one-line guard on the Compare: dropping inner.lineno == inner.end_lineno == node.lineno survives	Fixed in fae6b69: a comparison on another line is passed over, with a case and a mutant.
claude-review#3	minor	PR-3	tools/untested.py:499	No test holds the SWAP guard on the comparison's operator: a mutant that removes it survives, and the change has no mutant for it	Fixed in fae6b69 and c0065ce: what is read from a name an import binds is code.
claude-security#1	minor	PR-3	tools/untested.py:200	_literal still treats some world-reading entries in the mutant lists as data: any receiver's .replace, and subscripts or attributes of modules	Fixed in fae6b69, as claude-review#3 and agy#1.
claude-review-r2#1	major	PR-3	tools/untested.py:192	An import nested in if/try/with, or a star import, binds nothing in _bind, so what it gives is data in the mutant lists	Fixed in c0065ce: a name is data only where a statement before the list bound it; guarded and star imports and ENV = os.environ are code.
claude-review-r2#2	major	PR-3	tools/untested.py:210	No test holds the binding _bind reads from statements other than def/import/assign (for, with, AugAssign, class)	Fixed in c0065ce: a for loop's rebinding is held, with a case and a mutant.
claude-review-r2#3	minor	PR-3	tools/untested.py:193	No test holds that a dotted import binds its first name as code	Fixed in c0065ce: a dotted import binds its first name, with a case and a mutant.
claude-review-r2#4	minor	PR-3	tools/untested.py:207	No test holds that an unpacking other than A, B = x, y binds its names as code	Fixed in c0065ce: another unpacking binds its names as code, with a case and a mutant.
claude-review-r3#1	major	PR-3	tools/untested.py:193	A name rebound or filled by something other than a plain assignment target keeps its first kind, so an entry that reads the environment is still data	Fixed in 2424df1: a walrus, a global, a method and globals() or eval unsettle a name, each with a case and a mutant.
claude-review-r3#2	minor	PR-3	tools/untested.py:193	No test notices that an annotated assignment is read as an assignment	Fixed in 2424df1: a dunder attribute and a decorated def are code.
claude-review-r3#3	minor	PR-3	tools/untested.py:231	An attribute chain off a helper or a text reaches os without a call, and is data	Fixed in 2424df1: replace on a helper's return is data only where the helper says -> str.
claude-review-r3#4	minor	PR-3	tools/untested.py:253	_text takes any helper's return for a text, so replace on a helper's Path is data while ROOT.replace is code	Fixed in 2424df1: an annotated text is held, with a mutant.
agy-r4#1	major	PR-3	tests/test_untested.py:354	Import and rebinding cases in test_a_call_in_the_mutant_lists_is_data_only_of_the_modules_own_helpers do not test what they claim	Declined: the bare import cases are vacuous now that an unbound name is code, but their bound-first siblings (os = 'a' then an import) kill each import mutant.
agy-r4#2	major	PR-3	tools/mutants.py:15389	Calls to helpers after globals() or eval() bypass bound["*"] == "code"	Fixed in 5626728: a helper's call is code after globals() or eval.
agy-r4#3	major	PR-3	tools/mutants.py:15406	Mutant "a helper defined after the mutant lists is theirs" disables binding before the lists instead	Fixed in dd25df3: the mutant is renamed "no statement before the mutant lists binds a name", for what it does.
agy-r4#4	major	PR-3	tools/untested.py:167	Listed chained assignment in data_lines skips _bind and leaves prior bound names settled	Declined: data_lines binds a listed statement's other names after it reads the list (the _bind after the loop), so x in MUTANTS = x = [...] is code for a later list.
agy-r4#5	major	PR-3	tools/untested.py:201	_bind pairs tuple unpacking 1-to-1 even when the RHS tuple contains starred items	Fixed in dd25df3: an unpacking with a starred item is no pairing, with a case and a mutant.
agy-r4#6	major	PR-3	tools/untested.py:214	_bind and _unsettle ignore del statements and in-place subscript mutations in functions	Fixed in dd25df3: del, and a store through a subscript or attribute, make a name code, each with a case and a mutant.
agy-r4#7	major	PR-3	tools/untested.py:247	_unsettle misses method mutations when a function precedes the variable or shares a nested name	Fixed in dd25df3 for a helper defined before its name is bound. Left: a nested scope's own name masking the module's, which only a contrived helper meets.
agy-r4#8	major	PR-3	tools/untested.py:273	_literal allows calling a helper after globals() or exec() unsettles all names	Fixed in 5626728, as agy-r4#2.
claude-review-r4#1	minor	PR-3	tools/untested.py:286	_text's check that a call is itself literal is held by no test: a mutant that drops it survives	Fixed in 5626728: the literal check on a call read as a text is held, with a case and a mutant.
claude-review-r4#2	info	none	tools/untested.py:224	Most of REFLECTS, and the Store-local half of own, are held by no test	Fixed in 5626728, as agy-r4#2.
claude-review-r4#3	info	none	tools/untested.py:247	A helper that changes a module list before the list is bound, or through a subscript, leaves the name data	Fixed in dd25df3 (a helper filling a list early, a subscript store), as agy-r4#6 and #7.
claude-review-r4#4	info	none	tools/untested.py:273	After a call of globals() or eval, a helper's call is still data, though decision 0166 says such a name is code	Left: exec, vars, locals, setattr and delattr share globals' path, which globals() and eval hold; the Store half of own only makes more entries code.
claude-review-r4#5	info	none	tools/untested.py:406	Text methods other than replace, and replace on a helper's tuple item, are now code where main took them for data	Deliberate: only replace on a text is data; no entry of the real lists changes.

🤖 Generated with Claude Code