test(tokenizers): end-to-end Qwen3.8-27B template render with a tool conversation — the composition of #410/#411/#415 is unverified

[jamesburton]

Blocked until #411 and #415 are both merged. Small, but it closes a claim that is currently unverified in-repo.

The gap

Three PRs together make the real Qwen3.8-27B chat template usable:

PR	Adds
#410	qwen35 / qwen2 pre-tokenizer entries
#411	loop.previtem / loop.nextitem, is undefined
#415	parenthesized tuple literals

#415's author verified — via a local, uncommitted patch that no longer exists — that once #411 lands, the full unmodified template renders end-to-end with a tool-call/tool-response conversation and no further errors. Its independent reviewer flagged, correctly, that this rests entirely on the author's word and is reproducible by nothing in the repo.

Concretely: the committed FullTemplate_Render_KnownRemainingGapIsTracked test uses a single plain user message, so even once #411 lands and it starts succeeding, it never exercises the template's message.role == "tool" branches (fixture lines 148-154) — which are exactly the branches that consume loop.previtem / loop.nextitem, i.e. the whole reason #399/#411 existed.

So the headline claim "Qwen3.8-27B chat templating works" would be true and untested.

Ask

• An end-to-end test rendering the full unmodified Qwen/Qwen3.8-27B chat_template.jinja over a conversation containing: a system turn, a user turn, an assistant turn with tool calls, and consecutive tool-response messages (consecutive matters — that is what makes previtem/nextitem non-trivial).
• Assert the rendered output, not merely that it did not throw. A render that silently produces a malformed prompt is the failure mode that matters — see bug(tokenizers): Jinja attribute access on Undefined returns Undefined where Jinja2 raises — silently malformed prompts #416, where attribute access on Undefined returns Undefined instead of raising, so a broken template renders as an empty string.
• Per docs(process): require tests to be DEMONSTRATED to fail — discrimination is asserted but not proven #417, demonstrate the test can fail: revert any one of the three PRs' changes and confirm it does. Record which mutants were applied.
• Exercise reasoning_effort / enable_thinking paths if feat(tokenizers/server): surface reasoning_effort / enable_thinking chat-template knobs #400 has landed by then; otherwise note it as uncovered rather than implying coverage.

Why it is worth a ticket

The three PRs each verified their own piece honestly. Nothing verifies the composition, and composition is what the user actually gets. This is also the first concrete deliverable of the Qwen3.8-27B support work, so it should be provably true rather than inferred from three green PRs.

Related: #399, #409, #410, #411, #415, #416, #417. Context: .docs/model-target-qwen3.8-27b.md.