Declare the Anthropic intermediate-system contract for Qwen3.8 small

[fzyzcjy]

_common.run_one() has required anthropic_intermediate_system_expectation for every Anthropic run since #2711. test_qwen38small.py, added by #2760, never declared it, so its stage-c-2-gpu-h200 lane raises before doing any work:

File ".../test_session_server_multi_role/_common.py", line 81, in run_one
    raise ValueError("Anthropic per-model verification requires an intermediate-system expectation")

#2760 branched before that check existed, and its own CI had stage-c-2-gpu-h200 skipped, so the combination was never exercised. Observed on the current main commit 6b7ac37283 in https://github.com/radixark/miles/actions/runs/33730559131.

"forbidden" is derived, not guessed. _verify_intermediate_system() returns route_supports and "system" in fixed_template_append_roles(tito_model), and Qwen38SmallTITOTokenizer.FIXED_TEMPLATE.allowed_append_roles is {"tool", "user", "assistant"}, so the live capability is always False and only "forbidden" satisfies _assert_intermediate_system_expectation. The sibling models whose fixed templates also omit system, qwen35 and qwen36, declare "forbidden" as well; the ones that allow appending system declare "required".

Nothing enumerated the real ModelConfig objects, which is why a missing required field could only surface from a GPU lane. The second commit hunk adds a CPU regression test that parses every model config in that directory and asserts each Anthropic-verified one declares a valid contract. It reads the files with ast rather than importing them, so it does not trigger the e2e CI registrations. Verified both ways on a devbox: 9/9 parametrized cases pass as committed, and dropping the new line again fails exactly [test_qwen38small].

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[fzyzcjy]

Evidence that main is red without this change

1. Reproduced on pristine main, no GPU needed

run_one() raises before any engine is launched, so the break is reachable from a plain CPU box.
Checked out main at 6b7ac37283 with zero local modifications
(git diff origin/main --stat -- tests/e2e/sglang/test_session_server_multi_role/ is empty):

$ python -c "
from tests.e2e.sglang.test_session_server_multi_role import test_qwen38small as m
from tests.e2e.sglang.test_session_server_multi_role._common import run_one
print('tito_model =', m.CONFIG.tito_model)
print('verify_anthropic =', m.CONFIG.verify_anthropic)
print('anthropic_intermediate_system_expectation =', m.CONFIG.anthropic_intermediate_system_expectation)
run_one(m.CONFIG, session_server_version='v2', endpoint='anthropic')
"
tito_model = qwen38small
verify_anthropic = True
anthropic_intermediate_system_expectation = None
Traceback (most recent call last):
  File "<string>", line 7, in <module>
  File "/root/miles/tests/e2e/sglang/test_session_server_multi_role/_common.py", line 81, in run_one
    raise ValueError("Anthropic per-model verification requires an intermediate-system expectation")
ValueError: Anthropic per-model verification requires an intermediate-system expectation

run_both_versions() drives (v1, openai), (v2, openai), (v2, anthropic), and
verify_anthropic is True, so the third leg is always reached and always raises.

2. Observed as a real CI failure on this exact main content

Run 33730559131, job stage-c-2-gpu-h200 (0):

File ".../test_session_server_multi_role/_common.py", line 81, in run_one
    raise ValueError("Anthropic per-model verification requires an intermediate-system expectation")
FAILED: tests/e2e/sglang/test_session_server_multi_role/test_qwen38small.py returned exit code 1

That run is on a branch based on 6b7ac37283, and at that commit
git diff 6b7ac37283 HEAD -- tests/e2e/sglang/test_session_server_multi_role/ was empty — the whole
directory was byte-identical to main, so the branch contributed nothing to the failure.

3. Why nobody noticed

_common.py has required the field since #2711 (merged 2026-08-29). test_qwen38small.py arrived with
#2760 (merged 2026-08-31), whose branch head aff38c14d7 predates that check — its _common.py had no
such requirement, so the conflict was invisible to both sides. And #2760's own CI
(run 33139416968) has
stage-c-2-gpu-h200 skipped, so this file has never once run green.

4. The regression test in this PR does catch it

Verified both directions on a devbox against this branch:

• as committed: pytest tests/fast/utils/test_utils/test_session_verify_runner.py → 49 passed,
  including all 9 parametrized cases of the new test;
• with the one-line fix removed again: 1 failed, 8 passed, failing exactly
  test_every_anthropic_verified_model_declares_its_intermediate_system_contract[test_qwen38small].

So the CPU lane would have caught this at #2760 time, without waiting for a GPU lane that was skipped.

[fzyzcjy]

Second commit: the lane needs a second field too

Pushed d0af6543ee. Context, since it shows how thoroughly this file was never exercised.

Once the intermediate-system contract is declared, the lane gets past run_one() and actually reaches
the Anthropic leg for the first time ever. It then fails on the next gate — observed on
run 33746767704, job stage-c-2-gpu-h200 (0):

AssertionError: Session multi-role e2e: assistant_text mismatch ratio 32/32=1.000 exceeds threshold 0.2.

The metrics from that same run show every hard gate at zero:

'rollout/tito_session_mismatch_rate/v2/special_token_count': 0.0
'rollout/tito_session_mismatch_rate/v2/special_token_type':  0.0
'rollout/tito_session_mismatch_rate/v2/non_assistant_text':  0.0
'rollout/tito_session_mismatch_rate/v2/assistant_text':      1.0

Only the soft assistant_text tier moved, which is the endpoint-only serialization difference the
Anthropic tool-call conversion introduces — not a TITO bug.

This is a solved problem for this exact model family. Every sibling that pairs
tool_call_parser="qwen3_coder" with the Anthropic endpoint already carries the same override:

model	tool_call_parser	anthropic_assistant_text_threshold
qwen3	qwen25	1.0
qwen35	qwen3_coder	1.0
qwen36	qwen3_coder	1.0
qwen38small	qwen3_coder	missing → this commit

So qwen38small gets the same three lines qwen36 has, comment included. The OpenAI-side
assistant_text_threshold is untouched and every hard gate stays at 0, so this is not a weakened test —
it is the per-model Anthropic override the other three already use.