[SaladDay]

Why

Catch performance regressions before merge while keeping ordinary PR checks short. Long-history testing runs only when requested with perf:full or manual dispatch.

Changes

Add a performance workflow based on a pinned harness-perf-benchmark. It compares the PR base and head on the same macOS runner using an offline mock, measures duration, CPU and peak RSS, and saves machine configuration and raw logs.

Update CONTRIBUTING.md and the PR template: performance optimizations, large runtime refactors, and changes to history, token estimation, storage, streaming or tool execution require perf:full.

Behavior

Trigger	Behavior
Ordinary PR	Run the 100-round basic suite
Add perf:full	Run startup, 100-round and 300-round long-history scenarios
Push or reopen while labeled	Rerun the full suite
Remove perf:full	Return to basic and cancel the previous automatic run
Change another label	Leave performance runs untouched
Manual dispatch	Choose basic/full and the baseline

The check is named performance. Confirmed regressions, incorrect execution and invalid or excessively noisy measurements fail the check. Basic runs have a 15-minute limit; full runs have a 45-minute limit.