Running this contamination gate on git diffs zeroes out its literal check

The literal half of the score is the half that catches a shared sentinel UUID, and under the invocation the post prints, it cannot fire. `literals_from_ast()` parses the text it is handed with `ast.parse` and returns an empty set when that raises `SyntaxError` ^[9]. What it gets handed is `/tmp/prod.diff` and `/tmp/test.diff`, written by `git diff --unified=0 origin/main...HEAD` ^[10]. Diff text is not Python. Both literal sets come back empty, the intersection is empty, and `lit_score` evaluates to 0 over `max(1, 0)`, which is 0.0 ^[1]. Nothing reaches the 0.25 branch of the contamination test ^[7]. The report prints `literal_overlap: 0.0` and an empty `shared_literals` list. That output is indistinguishable from an uncontaminated pair.

The token side does survive the diff, because `tokens()` runs three regexes over raw text: identifiers of three characters or more, quoted strings of three or more, numbers of two digits or more, minus a fourteen-word stop list ^[12]. Both files handed to the gate are output from the same command, so whatever vocabulary the diff format contributes to each side lands in the intersection and counts as shared ^[2]. The denominator is the test's token count ^[6]. Scaffolding in the test diff lowers the score; vocabulary shared by both sides raises it.

Reading the two files instead of the two diffs fixes the AST and reintroduces the problem the author was avoiding. Whole-repo token sets are dominated by project vocabulary, and the post says that inflates the score and hides the leak ^[11]. At file scope the dilution is milder than at repo scope, but it is still imports, test framework names, and the public names the post tells you to expect, `compute_fee` and `FeeRequest` ^[13].

The second failure pattern was never going to show up in either score. "The production function grows an extra branch that exists only to satisfy a comment in the test," the post says, and coverage rises while behavior for real callers does not ^[4]. For that one it prescribes literal mutation plus an oracle session that never sees the patch ^[5]. The mutation helper appends `_x` to string constants of four characters or more and adds 1 to integers with absolute value of 10 or more ^[15]; if the suite still passes, the assertion was never bound to that value ^[16]. Constants outside those guards go through untouched, so an assertion pinned to 0, to 1, or to a three-character status string is not exercised by it ^[3].

Rewriting the oracle in a session that never sees the patch is the one prescription the post makes and does not implement ^[17]. The post is explicit about the status of the code that is there: a proposed local gate, not a published benchmark, not run against a private corpus, with the instruction "Label it as a method, not a result" ^[14]. The mutation helper carries the same warning in its header comment ^[18].