test(e2e): verify multi-turn Anthropic sessions

[guapisolo]

Summary

Verify six-request Anthropic sessions and pin SGLang conversion behavior in Miles CI.

Motivation

Six-request tool/result/text histories, reasoning replay, and supported intermediate-system insertion must preserve TITO prefixes. Incomplete Nemotron and GLM-4.7 responses must retry without advancing invalid history. SGLang PR #35127 refactors the runtime-independent Anthropic conversion seams, but its external-PR policy does not run the owning CPU suites, so Miles needs a temporary paired-checkout gate for behavior preservation.

Usage

python3 tests/e2e/sglang/test_session_server_multi_role/test_qwen36.py
SGLANG_SOURCE_ROOT=/path/to/sglang/python python3 -m pytest -q tests/fast/router/test_sglang_anthropic_conversion.py

Design Notes

• Mental model: Each per-model test drives OpenAI v1/v2 and Anthropic v2 through the session tree.
• Verification gates: The verifier admits only a valid six-request leaf, then independently gates hard token drift and model/endpoint-specific assistant serialization drift.
• CPU ownership: Miles runs SGLang's conversion tests from SGLANG_SOURCE_ROOT pinned by ci-sglang-pr: #35127 at e95eae48.
• Retries: Invalid responses retry from the same parent.
• Intermediate system: Insertion requires both live-route and template support.
• MiniMax: The test skips Anthropic and rejects direct agent use.
• Port isolation: Auto allocation selects each worker port independently; an explicit --session-server-port still anchors a consecutive range.
• Formatting boundary: Qwen3/Qwen3.5/Qwen3.6 keep Anthropic-only 1.0 assistant-text thresholds for conversion-normalized delimiters; OpenAI thresholds and all hard mismatch gates remain unchanged.

Verification

• Exact-image focused tests on miles-h200-dev-2: 38 passed for session verifier routing on b9bea31d, then 12 passed for router allocation and 146 passed for argument parsing on 713df081c; changed-file pre-commit passed.
• Qwen3 full v1/v2/Anthropic E2E on two H200s: two consecutive runs passed.
• Qwen3.6 Anthropic on four H200s: four samples completed six requests each; hard mismatch metrics were 0.0, and the captured assistant diff was only </think>\n\n<tool_call> versus </think><tool_call>.
• Controlled Qwen3.6 gate: the captured 4/4 sidecar failed at 0.2 and passed at the Anthropic-only 1.0; pre-commit passed.
• Rebase at b8f5667c: 11 commits replayed conflict-free onto 11563829a; source metadata and cumulative trees were verified.
• git diff --check 11563829a...b8f5667c: passed.

Review Focus

• Anthropic verifier: canonical-record assertions, retry semantics, and hard-versus-soft mismatch boundaries.
• CPU bridge: paired SGLang checkout at e95eae48 and subprocess failure propagation.
• Port allocator: independent automatic allocation and explicit-range preservation.

[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.}

[guapisolo]

Wait for sglang bump and rebase, then we can merge this PR.

[guapisolo]

Verified on head 18d41fd0806d39adccc77d3b49631926151ef6f1:

• Run pre-commit passed after accepting the repository's Black formatting.
• stage-a-cpu (2) / run-cpu selected and passed test_sglang_anthropic_conversion_contract (18.82s).

The bridge executes SGLang #35127's owned test_utils.py and test_serving.py; the direct paired-checkout run on miles-h200-dev-2 reported 84 passed, 55 subtests passed.