docs(parsers): add DIS-1926 vLLM tool-parser test audit

[zhongdaor-nv]

Overview

Companion artifact to #9290 (already merged — refined PARSER_CASES.md taxonomy from this audit). This PR adds the full per-test bidirectional audit that informed every change in #9290.

Marked draft because reviewers may want to argue per-row bucket assignments; the audit is a working reference doc rather than a stable contract.

What's in this PR

lib/parsers/VLLM_TEST_AUDIT.md (new file, 906 lines, 493 distinct test rows):

• Source: vLLM main at b53c507bc91f87e28b03e9b54bbff7c76e97d58b (vllm/tool_parsers/*, tests/tool_parsers/*, tests/tool_use/*, tests/entrypoints/openai/tool_parsers/*).
• Scope: 421 explicit test functions + 72 inherited common-suite rows from ToolParserTests.
• Bucketing: every row carries one or more PARSER_CASES.md / REASONING_CASES.md / PIPELINE_CASES.md / FRONTEND_CASES.md tags, plus a clickable GitHub source link, plus a one-line behavioral note.

Re-bucketing summary (post PR #9127 taxonomy)

The audit was originally written against the old CASE.* labels in lib/parsers/TEST_CASES.md (deleted by #9127). Mechanical renames + per-row classification done:

• 244 streaming rows split per-row into PARSER.stream.{1,2,3,4} (single-call assembly / multi-call assembly / partial-token chunking / streaming termination)
• 26 fmt rows split per-row into PARSER.fmt.1 (function-name surface) vs PARSER.fmt.5 (argument-envelope shape: native ID, JSON field-order, arguments ↔ parameters alias)
• Out-of-PARSER-scope buckets relocated: CASE.{11,18,25} → FRONTEND.{1,3,5,6}; CASE.12 → PIPELINE.finish_reason; CASE.{9,10,17} → REASONING.batch.{1,2}; CASE.20 → // helper; CASE.16 → inline-regression annotation; CASE.26 dissolved into PARSER.batch.4 impl-defined recovery contract

Mis-bucket fixes caught by review

• FunctionGemma::test_multiple_tool_calls and Gemma4::TestExtractToolCalls.test_multiple_tool_calls were both labeled CASE.1 but assert len(tool_calls) == 2 — corrected to PARSER.batch.2.

Bucket-assignment refinements caught by review

• test_unique_tool_call_ids (DSv3.2): drops fmt.5 (no native call-ID surface)
• test_invalid_funcall_id_skipped (Kimi K2): fmt.5 → fmt.1 (validation, not preservation)
• 3 Mistral argument_before_name* parametrized rows: added missing fmt.5 tag

A staleness banner at the top documents the re-bucketing transformation and mis-bucket fixes for traceability.

Top findings (already in PR #9290 or flagged for follow-up)

1. Mistral v11+ wire format — STILL OPEN (parser doesn't exist; flagged in PARSER_CASES.md "Known production gaps")
2. PARSER.stream.{1..4} parser-tier in Kimi K2 / Qwen3 / Hermes / Pythonic / Mistral — partial closure via DSv4 (test(parsers): DIS-1842 — DSv4 + Kimi K2 unit-test coverage gaps #8946) + Gemma 4 (chore(frontend): Add Gemma 4 parser support + Test Cases #8852); 5 families remain
3. FRONTEND.3 (adjust_request) — CLOSED for 7 families via 28 new tests in lib/llm/tests/tool_choice.rs (test(parsers): DIS-1842 — DSv4 + Kimi K2 unit-test coverage gaps #8946 + test(parsers): Top-N models to have extra CASE.6+ coverage (case3) #9035)

Test plan

• cargo check -p dynamo-parsers --tests passes (docs-only; new file)
• All 493 test rows carry at least one new-taxonomy bucket
• Bucket-summary tally adds up; "Total (distinct test rows) = 493" matches header
• No leftover old CASE.* labels outside the staleness banner / "Old label" column / mis-bucket annotation notes
• Internal review caught 4 bucket-assignment issues + 2 mis-bucketings; all fixed

Out of scope / follow-ups

• Per-family ticket spawning: the 5 PARSER.stream.* gaps and the Mistral v11 parser are separate work items
• DIS-1906 (cross-impl parser parity harness) can use this audit as fixture-row source

---

Closes the audit half of DIS-1926 (the taxonomy half landed in #9290).

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-f6c88b3d-9658-488b-9af2-dc83b9126bca.docs.buildwithfern.com/dynamo/dev