@codex review

Codex Review: Didn't find any major issues. Already looking forward to the next diff.

Reviewed commit: dd400ba8ed

@codex review

Codex Review: Didn't find any major issues. Swish!

Reviewed commit: 8dd8ba68ba

@codex review

Codex Review: Didn't find any major issues. Keep them coming!

Reviewed commit: f64e498a54

Independent merge-worthiness check on this branch, since a characterization test is only
worth its line count if something can actually make it red that nothing else catches.

The two mutants in the description are already covered

Both rows in the table above do turn the new test red, but neither is a gap the branch
closes. Running each mutant against the merge base, restricted to the placeholder test
files:

The first two are caught for cost, not for meaning: both add an unbounded backward scan,
so the linearity test sees them before the semantics ever come up.

The mutant that this test, and only this test, catches

The "leave the oversized fragment whole" reading is implementable without giving up the
performance contract: look back a bounded distance for an opener that predates the
window, and decline if one is there. It stays O(1) in reply length, so the linearity test
cannot see it, and the long-reply cases carry no out-of-window ${, so those cannot
either.

if (from > 0) {
  const back = text.slice(Math.max(0, from - 2 * window), from);
  if (back.includes("${")) {
    return text;
  }
}
const tail = text.slice(from);

Full frontend suite, that mutant applied:

So it is the only guard in the suite against that change. That is the row the description
should be making.

Per assertion

The 4,208 assertion reads only the reference regex defined in the test file, so no change
to the module can move it, and weakening the reference regex is caught by five older tests
in the same file first. It documents the comparison rather than guarding anything. Worth
keeping for the reader, not worth counting as coverage.

Why the input class is not an enumeration artefact

The shape is reachable from ordinary prose. A reply that mentions an unterminated
expansion early and ends with the leaked placeholder: an opening line
"Sure. In bash, ${VAR" with no closing brace, 120 sentences of explanation, a short fenced bash block, then the provider's trailing ${answer}`.

8,370 characters. The unbounded pattern removes 8,354 of them and leaves
Sure. In bash, + backtick. The bounded scan removes 10, the newline and ${answer}.
That is the bug #9012 fixed, and this test pins the boundary that separates the fix from
a regression back to it.

Other checks

• studio/frontend/src tree hash is identical between the merge base and this head, and
  studio/backend differs only under tests/. No shipped byte changes, so old installs,
  existing threads, existing settings and per-model config cannot be affected.
• The three Python files are exactly what scripts/run_ruff_format.py produces when run
  on the merge base. Byte identical. Pre-existing drift on main, not this branch fighting
  the formatter.
• 100,000 generated inputs at the real 4,096 window, plus a 20,000 reply end-of-stream
  model and a 20,000 case interrupt/reload/second-tab model: no input where the bounded
  scan removed more than the unbounded pattern, returned a non-prefix, or where
  isCandidate() was false while the strip cut. 8,333 of the 20,000 replies were actually
  stripped, so the properties are not vacuous.
• Edge cases all agree with the unbounded pattern: empty, one character, ${},
  unterminated ${, 3,000 nested openers, unterminated code fence, RTL, CJK, emoji inside
  the fragment, a 300,000 character reply, and a 9,000 character trailing whitespace run.
• Chromium 151 / Firefox 153 / WebKit 26.5: identical results on the pinned case and on
  20,000 generated inputs each. The module is pure string and regex work, so the engines
  are what matter here, not the renderers.
• npm test 3,672 passing, 0 failing. npm run typecheck clean.
• Cross platform on a staging replica: frontend unit tests pass on ubuntu-latest and
  windows-latest, 3,672 of 3,672 on both, with the new test named in each log.

CI note: both reds are inherited from main, and neither is reachable from this branch, which changes only test files.

• (Python 3.13): the lan_access.py encoding failure main carries. Fixed in open PR Name the encoding when reading wslinfo's output #9142.
• GGUF inference smoke (API, tools, vision): red on every main run since studio: refuse a gguf that cannot fit in free vram plus available ram #8883, at the JSON schema decoding + image input step. Fixed in open PR Unblock Mac Studio GGUF CI, red on every main run since #8883 #9155.

Worth restating plainly for whoever merges this: git rev-parse gives the identical studio/frontend/src tree hash on the merge base and on this head, and all four changed files are tests. This PR ships no runtime bytes at all, so it cannot affect either job.