fix(parsing): schema-aware tool-arg coercion for XML-style parsers

[hallerite]

Summary

XML-style chat templates (Qwen3.5, GLM-5/4.5, MiniMax-M2, Laguna) render tool-call argument values verbatim inside <arg_value>X</arg_value> (or <parameter>X</parameter>) tags with no quoting. A value of true and the string "true" produce identical wire bytes — without the tool schema, the parser has no signal to choose between them and defaults to json.loads, silently corrupting string args that happen to look like JSON. This PR plumbs the schema in and preserves the type.

Raised by Robin (Poolside) on #21.

Changes

API change — additive, backwards-compatible:

class Renderer(Protocol):
    def parse_response(
        self,
        token_ids: list[int],
        *,
        tools: list[ToolSpec] | None = None,   # ← new
    ) -> ParsedResponse: ...

When tools=None (the historical default), behavior is unchanged. When tools is supplied, the four XML-style parsers consult each parameter's declared JSON-schema type and preserve declared-string params verbatim. Matches vLLM / SGLang reference parsers (glm45_tool_parser.py, hermes_tool_parser.py).

Helpers added to renderers/parsing.py:

• _build_param_type_index(tools) — accepts either flat ToolSpec or OpenAI envelope {"type":"function","function":{…}}, returns {tool_name: {param_name: schema_fragment}}.
• _coerce_arg_value(text, schema) -> (value, used_json_fallback) — declared-string → text verbatim; anything else → try json.loads, fall back to text. The bool lets callers flag INVALID_JSON only on genuine parse failures (not on schema-driven string preservation).

Parsers updated (with tools plumbing): parse_qwen35, parse_glm, parse_minimax, parse_laguna_xs2.

Parsers that accept-but-ignore tools (their wire formats quote strings, so schema isn't needed): Qwen3 hermes, Qwen3-VL, DeepSeek-V3, Kimi K2, Kimi K2.5, Nemotron3, gpt-oss harmony, Default.

Client (renderers/client.py): generate() already accepted tools; now forwards it through to parse_response so HTTP callers get schema-aware parsing automatically.

Downstream callers

Anything outside this repo that calls renderer.parse_response(token_ids) needs a one-line update to opt into schema-aware parsing:

parsed = renderer.parse_response(completion_ids, tools=tools)

Existing callers keep working unchanged (default tools=None preserves the historical behavior).

Test outcome

Model	Parser shape	Before	After
Qwen/Qwen3-8B	hermes JSON	✅ 5/5	✅ 5/5
moonshotai/Kimi-K2-Instruct	section JSON	✅ 5/5	✅ 5/5
Qwen/Qwen3.5-9B	XML	❌ 0/5	✅ 5/5
zai-org/GLM-5	XML	❌ 0/5	✅ 5/5
MiniMaxAI/MiniMax-M2.5	XML	❌ 0/5	✅ 5/5
poolside/Laguna-XS.2	XML	❌ 0/5	✅ 5/5

Test plan

• uv run pytest tests/test_tool_arg_type_preservation.py → 30 passed
• uv run pytest tests/ (full suite) → 1067 passed, 49 skipped, 1 xfailed, 0 failed
• uv run ruff check renderers/ tests/ clean
• uv run ruff format --check renderers/ tests/ clean

🤖 Generated with Claude Code