chat : enforce string enums in Qwen XML tool arguments

[arc-uri-el]

Overview

The Qwen XML tool grammar accepts arbitrary strings for a parameter declared as {"type":"string","enum":["list","load"]}. For example, <parameter=action>\nerase\n</parameter> is accepted even though erase is not an allowed value.

The handler selects its unrestricted XML string rule whenever the parameter resolves to a string. It therefore bypasses the JSON converter that would enforce the enum for JSON arguments.

For a nonempty string enum, build the XML value rule from the allowed literal strings instead. Check the closing-parameter boundary when choosing each literal, so an earlier shorter entry does not consume the prefix of a longer one. Ordinary strings and non-string enums retain their existing paths. This patch adds no root-union support, schema adapter, public API, model or GPU change.

Additional information

Validation

Base: 33870a2e1c99419041ae57f838db0ba3f9d6a122.

Separate matching baseline/candidate Linux x86_64 builds used GCC 15.2, Release, shared libraries, native CPU and OpenMP. All seven selected parser/grammar suites passed on both builds, with zero compiler warnings: test-chat, test-json-schema-to-grammar, test-chat-peg-parser, test-chat-auto-parser, test-chat-template, test-grammar-parser and test-grammar-integration.

The exact new test-chat binary fails against baseline libraries at the invalid-enum assertion (expected rejection, observed acceptance) and passes against candidate libraries. The added test covers Qwen3-Coder and Qwen3.5 templates, invalid strings, shared-prefix alternatives, empty strings, embedded quotes/newlines, parsed/streamed tool arguments, parallel calls and an unrestricted-string positive control.

This replaces the broader root-union proposal in #25. That representation can be adapted on the client side; ordinary enum enforcement is an independent defect.

Not run for this isolated patch: model inference, full CI, other-platform builds, perplexity or performance sweeps. No throughput or general model-quality improvement is claimed.

Requirements

• I have read and agree with the contributing guidelines.
• This is a human-directed contribution. I own its review follow-up.
• AI usage disclosure: YES. GPT-6 Astra assisted with investigation, code and regression tests, execution, evidence analysis and submission preparation. No independent human code-review attestation is claimed.

[dzannotti]

@arc-uri-el LGTM, but it has conflicts, would you mind rebasing?

[arc-uri-el]

Rebased onto current halo-box:master and force-pushed the updated branch. The parser change is now in common/parsers/qwen3-coder.cpp after the upstream parser split. A fresh Release shared-library build of test-chat with GCC 15.2 completed, and the full test-chat suite passed.