Agent Timeline local-media benchmark runner

runAgentTimelineLocalBenchmark is a development-bridge-only tool for scripts/agent-timeline/collect-real-media-benchmark.mjs. It is deliberately absent from the normal AI/chat tool registry. Invoke it through the usual authenticated /api/ai-tools bridge; the collector can select a fresh tab automatically or target one explicitly with --target-tab-id, and standard bridge tab isolation applies unchanged.

The runner accepts the collector’s local baseline/analysis request schema and the separate { "cancel": true } request. It never reads the supplied path from browser JavaScript. Instead, it requires exactly one selected local Media panel item whose browser File name and size match the supplied local fingerprint. Visual/audio passes additionally require one matching timeline clip. The SHA-256 remains collector evidence and is not recomputed or trusted as browser-local file access.

Supported local adapters are scene cuts, ClipAnalyzer metrics (focus/motion), ClipAnalyzer faces, and the existing local audio-level analyzer. They do not invoke transcription, scene descriptions, cloud APIs, or network-backed model providers. A second request with { "cancel": true } asks supported active local analyzers to cancel; audio decoding has no cancellable adapter and is reported without a fabricated cancellation claim.

The runner has two separate contracts: a baseline pass and an analysis pass. Both echo the requested baselineKind (standalone-cut or proxy-piggyback), and the collector rejects a result whose pass, baseline kind, platform, device class, or renderer/backend evidence differs from its paired result. The current browser adapter can execute a standalone-cut baseline, but it is a completed measurement only when its requested cache state is verified and an explicit local instrumentation capability is registered; proxy-piggyback has no instrumented adapter. This repository defines those capability hooks but does not register them, so the browser runner cannot emit a qualifying real-media measurement.

Peak memory, durable artifact bytes, redundant decoded seconds, and optional renderer/backend/software-fallback evidence are injected only by a local instrumentation capability that observed that exact pass. The runner never substitutes zero. Without the capability it fails closed as unavailable, so the collector cannot produce a qualifying performance claim.

Cold reset is not asserted from an empty UI state. A registered local verifier must directly confirm analyzer/model/artifact reset before a cold pass runs. Warm cache is only labelled warm when matching local output is visible and the pass reports measured redundantDecodedSeconds: 0; absent or non-zero decoder evidence is blocked. Linux/Mesa policies can require observable platform class, Mesa state, renderer backend, and software-canvas fallback.

npm run agent-timeline:benchmark also runs the separate deterministic synthetic benchmark. Its reports validate benchmark contracts only; they never qualify for the production analysis benchmark gate.