[SiluPanda]

Summary

• require end-to-end latency in new preset-agent benchmark reports and derive per-user decode throughput from E2E latency, TTFT, and measured output-token totals
• stop interpreting SGLang's chunk-level ITL as TPOT by allowing unavailable TPOT values to be reported as null
• use the derived rate in both stored preset rows and in-progress session rows, while retaining a TPOT fallback for presets and sessions created by older clients
• reject fresh reports whose claimed TPOT would require more decode time than the measured duration and concurrency can supply

The agent prompt now calls out SGLang's speculative-decoding behavior and explains which latency fields to record. At the storage boundary, the stricter fresh-report model is normalized back to the backward-compatible persisted model.

Fixes #4225.

Validation

$ uv run pytest -q src/tests/_internal/cli/services/presets src/tests/_internal/cli/models/test_presets.py src/tests/_internal/core/models/test_presets.py
305 passed, 2 skipped in 3.30s

$ uv run ruff format --check <changed Python files>
11 files already formatted

$ uv run ruff check <changed Python files>
All checks passed!

$ git diff --check
# no output

AI-assisted implementation; I reviewed the diff and ran the checks above locally.