fix: support DeepSeek-V4-Flash-0731 reasoning effort

[ormandj]

Motivation

DeepSeek-V4-Flash-0731/encoding/encoding_dsv4.py defines three reasoning-effort tiers. SGLang currently preserves the earlier DeepSeek-V4 contract, where high adds no prefix and max adds the "Absolute maximum" prefix. Applying the newer mapping to every DeepSeek-V4 checkpoint would change existing behavior.

Checkpoint family	low	high	max
Earlier DeepSeek-V4	filtered to the default	no prefix	"Absolute maximum" prefix
DeepSeek-V4-Flash-0731	no prefix	"Absolute maximum" prefix	"Beyond maximum" prefix

Modifications

• Select the newer mapping with SGLang's existing checkpoint_bundles_dspark_draft() helper. The HF config carries no encoding-version field, so this uses the bundled dspark_* config keys as the proxy; any DeepSeek-V4 checkpoint carrying those keys selects the newer mapping.
• Add the 0731 low, high, and max prompts from the bundled encoder.
• Preserve the existing earlier-checkpoint mapping.
• Add unit coverage for checkpoint detection, serving-layer filtering, all three newer tiers, invalid values, and legacy compatibility.

Accuracy Tests

• uv run --no-project python -m pytest -q test/registered/unit/entrypoints/openai/test_serving_chat.py: 86 passed and 30 subtests passed.
• On a TP2 server using DeepSeek-V4-Flash-0731, a one-message "Hello" request with max_tokens=8 produced prompt-token counts of 5, 84, and 97 for low, high, and max, respectively.

Speed Tests and Profiling

Not run. This changes which prompt prefix is emitted.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations. The SGLANG_DSV4_REASONING_EFFORT reference now describes the checkpoint-specific accepted values.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed. No accuracy or speed benchmark was run; the focused prompt-encoding coverage and prompt-token observation are listed above.
• Follow the SGLang code style guidance.

AI was used to assist with implementation.

---

CI States

Latest PR Test (Base): ❌ Run #30674188412
Latest PR Test (Extra): ❌ Run #30674188407

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[ormandj]

@mmangkad @JustinTong0323 @b8zhong @AliceChenyy, could you review this alongside #33140? The functional distinction is checkpoint selection: this version uses the existing checkpoint_bundles_dspark_draft() config signal instead of a model-path substring. It also updates the SGLANG_DSV4_REASONING_EFFORT documentation and covers request filtering, invalid values, and legacy behavior.

[JustinTong0323]

Thanks @ormandj. The DeepSeek V4 reasoning-effort work from this PR has been merged into #33140, with co-author attribution preserved there. Closing this PR in favor of #33140.