fix: GLM 4.5 streaming tool-call parsing + grammar error handling

[Gunther-Schulz]

Disclosure: All code and this PR description were written by AI (Cursor).

---

Summary

This PR fixes GLM 4.5 (and related XML tool-call) streaming behaviour so that:

1. The server no longer hangs when tool_choice=auto and the model never outputs the grammar trigger (parse-only path, no grammar for AUTO).
2. The "Failed to parse up to error: ... attempting to parse an empty input" log spam is removed when streaming partial tool-call arguments.
3. Plain-text arg_value content (e.g. subagent_type=explore) is parsed correctly instead of throwing partial forever, so Task/tool calls complete.
4. Grammar/sampler runtime errors (e.g. "Unexpected empty grammar stack") return HTTP 500 and release the slot instead of aborting the server.

Changes

1. GLM 4.5 parse-only for AUTO (existing branch commit)

• common/chat.cpp (GLM 4.5 init): Use grammar only when tool_choice == required; for auto, do not set grammar/trigger so tool calls are detected by parsing decoded text (vLLM-style).
• Tests: Relax test-chat assert to allow empty grammar when the test message has tool_calls.
• Addresses server hang when the model never outputs the trigger (e.g. Misc. bug: GLM-4.7-Flash enters corrupted state with grammar trigger loop, memory leak, and gibberish output #19068).

2. Avoid "parse empty input" log (json-partial + XML parser)

• common/json-partial.cpp: When the SAX parser reports an error at position 0 (empty substring to re-parse), return false immediately without calling json::parse or logging "Failed to parse up to error". This avoids noisy logs for all formats (GLM 4.5, GPT-OSS, Granite, Hermes, etc.) when streaming partial or invalid JSON.
• common/chat-parser-xml-toolcall.cpp: Only call try_consume_json() when the remainder after <arg_value> looks like the start of a JSON value (", {, [, digit, -, or prefix of true/false/null). Otherwise treat as incomplete/plain text and skip the JSON parser (vLLM-style), avoiding SAX error-at-position-0 for e.g. exp before explore.

3. Plain-text arg_value fallthrough + server grammar error handling

• common/chat-parser-xml-toolcall.cpp: When the remainder does not look like JSON (e.g. explore), do not throw; skip try_consume_json() and fall through to the existing plain-text path (try_find_val_end()). Fixes Task tool (e.g. subagent_type=explore) never completing.
• tools/server/server-context.cpp: Wrap common_sampler_accept() and common_sampler_sample_and_accept_n() in try/catch for std::runtime_error. On catch (e.g. "Unexpected empty grammar stack" from llama-grammar.cpp), log, send HTTP 500 with "Grammar constraint violation", release the slot, and continue so the server stays up.

Testing

• test-json-partial and test-chat-parser pass.
• Manual: GLM 4.5 with tool_choice=auto, streaming tool calls (e.g. Task with subagent_type=explore) complete without log spam; server does not abort on grammar errors.

Related

• vLLM behaviour: GLM-4 parser only parses non-string arg values when </arg_value> is seen; we align with that.
• Fixes / improves experience around streaming XML tool calls and grammar constraint violations in the server.

[pwilkin]

Obsoleted by #18675