Add fix-tool-choice-enforcement mod

[tea24864]

What this fixes

On affected builds, tool_choice="required" and named tool choice on
/v1/chat/completions decode free-form. The model answers in prose and no tool
call is produced.

A required request is also malformed on the way out: finish_reason is
"tool_calls" while tool_calls is null, because the serving layer forces
that value and then finds nothing to report. Any OpenAI-compatible client that
branches on finish_reason breaks on it.

Reproduced on vllm-node-b12x builds 0.1.dev20489+ga50ebee1d.d20260904 and
0.1.dev20596+g2a979314d.d20260907, serving GLM-5.3-Flash with
--reasoning-parser glm45 --tool-call-parser glm47 --enable-auto-tool-choice.

Cause

Enforcement is an xgrammar structural tag built in Parser.adjust_request. Two
independent breaks keep it from reaching the sampler.

1. Nothing on the chat-completions path calls adjust_request. The only
   call site in vllm/entrypoints/ is the harmony branch of the Responses API,
   next to a TODO about unifying it with the non-harmony branch.
2. A collapsed parser engine never builds the tag.
   DelegatingParser.adjust_request is the only implementation that calls
   _apply_structural_tag. When ParserManager.get_parser collapses a
   reasoning and tool adapter that share one engine, which glm45 + glm47
   do, it returns the engine class itself, and that class has no structural-tag
   code at all.

Fixing either half alone changes nothing on a collapsing build. Part 1 was
applied on its own and both probes returned byte-identical output.

The second break is specific to the builds this mod targets. Stock
vllm-project/vllm has no _get_parser_engine_cls branch, so get_parser
always returns a DelegatingParser subclass and part 1 alone is enough there.
The mod README documents how to tell the two apart.

What the mod does

• Part 1, model-agnostic. Calls parser.adjust_request(request) in
  _create_chat_completion, after the parser is built and before the request is
  rendered, matching where the Responses API makes the call and where the tag
  has to land to reach to_sampling_params. Skipped when the installed vLLM
  already calls it.
• Part 2, GLM only. Gives Glm47MoeParser an adjust_request that keeps
  the engine's behaviour and then applies the tag, mirroring
  DelegatingParser._apply_structural_tag rather than inventing a second
  policy, including the VLLM_ENFORCE_STRICT_TOOL_CALLING check that
  AbstractToolParser.get_structural_tag applies. Skipped when
  vllm/parser/glm47_moe.py is not installed.

Blast radius

The tag builder returns None for tool_choice: "none", and for "auto"
unless a tool declares strict, so ordinary auto traffic decodes as before.
Confirmed live: an auto request still calls the tool when the question needs
one, still answers directly when it does not, and the reasoning block is intact
either way.

VLLM_ENFORCE_STRICT_TOOL_CALLING=0 disables the whole path, as upstream.

The mod is opt-in. No shipped recipe declares it.

Validation

4x DGX Spark TP4 cluster, tool-eval-bench v2.6.1 hardmode, 88 scenarios,
GLM-5.3-Flash EXL3 6bpw, 0.1.dev20489:

	points / 176	score	TC-45
Before	159	90	fail
After, trial 1	161	91	pass
After, trial 2	160	91	pass
After, trial 3	161	91	pass

TC-45 is tool_choice=required Compliance. Diffed scenario by scenario, ten
scenarios moved and the nine besides TC-45 cancel to exactly zero, so the net
gain is entirely this fix. EXL3 4bpw moved 154 to 158 points with the same flip.

Across 17 scored hardmode runs, TC-45 tracks the engine build rather than the
quantization or checkpoint, including a 0.1.dev20596 NVFP4-Spark run that
fails the same way and rules out the EXL3 checkpoints as the cause.

Tests

Adds tests/test_tool_choice_enforcement_mod.sh, following the existing mod
test convention. It covers call placement inside the parser_cls guard and
ahead of rendering, no leakage into the streaming path, a self-contained
appended block, idempotency, fail-closed on source drift in either file,
non-GLM images, an upstream that already applies the call, and a missing vLLM
install.

./tests/test_tool_choice_enforcement_mod.sh
./tests/test_recipes.sh -v                      # 67 passed
./tests/test_launch_cluster_image_sync.sh
./tests/test_launch_cluster_vllm_pr.sh
./tests/test_instanttensor_zero_copy_mod.sh     # regression
./tests/test_radixark_dspark_mod.sh             # regression

All top-level recipes still dry-run clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01A5YuMY6DErgC5h2Uztz93b

[tea24864]

Closing this: I opened it against the wrong repository. This was meant for my own fork, not upstream. Sorry for the noise.