[hetaoBackend]

Summary

The output-speed readout could spike at the very start of a turn. When a provider buffers upstream output and releases it in one burst, the decode window for that burst says nothing about model speed, but its tokens were still folded into the rate — so the first reported value was far above the model's real throughput.

A sample is now treated as a buffered burst when its decode window is under 10% of the request duration and its implied speed exceeds 1000 tok/s, and such samples are dropped instead of aggregated. Genuinely fast providers are unaffected: a 1900 tok/s sample spread over most of its request still counts.

Changes

• packages/tui/src/tui/controller/projection/turn-output-rate.ts — drop buffered burst samples before aggregation
• packages/tui/test/unit/tui/controller/projection/turn-output-rate.test.ts — cover burst, instant-burst, normal and fast-provider cases

Validation

• turn-output-rate.test.ts: 11 passed (1 new case)
• eslint on both changed files: clean
• tsc --noEmit -p tsconfig.standalone.json: clean
• Full capability gate: 5035 passed. The 340 failures in that gate are pre-existing and unrelated — an untouched main run in the same environment produces the identical 16 failed files / 340 failed tests, caused by a missing better-sqlite3 native binding in this sandbox.

Notes

This is a port of an internally released fix. The public distribution has no per-session subagent concurrency limit, so the two related changes shipped alongside it internally do not apply here.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}