Performance benchmarks: turn-pipeline latency targets on reference hardware

[charneykaye]

Establish and enforce latency targets for the turn pipeline on reference hardware (Apple M2, Intel i7 CPU-only, mid-spec Linux x86), per docs/performance.md. Requires real-model infrastructure — unblocked once #457 (real-model CI smoke) lands.

Scope

• Define budget per stage: input → prompt compose → inference first-token / full-turn → state update → render
• Scripted benchmark harness reusing the Automated real-model CI smoke test #457 nightly model cache; publish results as a CI artifact
• Regression alarm against the recorded baseline

Acceptance

• docs/performance.md gains a real, dated baseline table per reference machine
• Nightly job fails on >20% regression against baseline

From docs/post-alpha-issues.md item 8 (Milestone 2 — hardening). Blocked by #457.