Requirements

• I have added test coverage for new or changed functionality
• I have followed the repository's pull request submission guidelines
• I have validated my changes against all supported platform versions

Related issues

Addresses SEC-8020. Mirrors the fix applied in the Go server AI SDK (go-server-sdk commit 3317871). A parallel PR has also been opened for launchdarkly/js-core.

Describe the solution you've provided

The judge's _interpolate_message method previously used chevron.render() (Mustache templating) for its second-pass interpolation of {{message_history}} and {{response_to_evaluate}} placeholders. This is vulnerable to template injection: attacker-controlled values resolved during pass 1 (e.g. Mustache delimiter-change tags like {{=[ ]=}}) would be interpreted as control syntax by the second Mustache pass, potentially blinding the judge to the actual content being evaluated.

This PR replaces the Mustache-based interpolation with simple str.replace() calls. Since the judge only ever substitutes two known placeholder strings, a full template engine is unnecessary, and literal string replacement is both safer and simpler.

Note: chevron is still used in client.py for the first-pass template interpolation (which is expected behavior — pass 1 needs Mustache to resolve context variables).

Describe alternatives you've considered

• Escaping attacker-controlled values before passing them to Mustache for pass 2. This is fragile and hard to get right across all Mustache control sequences.
• Using a different template engine with sandboxing. Unnecessary complexity given only two fixed placeholders.

Additional context

Key items for review:

• Verify that the chevron import removal from the judge module doesn't affect other code paths (it doesn't — chevron is only used in client.py for pass 1).
• The regression tests cover all major Mustache injection vectors (delimiter changes, partials, comments, triple-stache, sections) plus edge cases like Mustache-like syntax appearing in the actual message content.

Updates since last revision

• Simplified the _interpolate_message docstring per reviewer feedback on the parallel JS PR — now a concise two-line summary instead of a verbose multi-line explanation.

Link to Devin session: https://app.devin.ai/sessions/651e799b906748a4834bafefb4a3e3e5
Requested by: @jsonbailey