perf(desktop): add a stream-history scenario to the perf harness

The existing `stream` scenario measures streaming into an empty thread, which
is exactly the case where per-delta transcript work does not show up. This adds
`--historyTurns`, which mounts a settled transcript and lets it drain BEFORE
the recorders start, so the measurement window contains streaming work only —
and `stream-history`, a report-only preset that turns it on.

Report-only (tier: manual) rather than gated: how much history a host can mount
varies, so the absolute number is not comparable across machines. It is meant
for same-machine before/after runs.

Co-authored-by: Jakub Wolniewicz <frizikk@users.noreply.github.com>

apps/desktop/scripts/perf/scenarios/index.mjs

@@ -8,11 +8,13 @@ import multitab from './multitab.mjs'
 import profileSwitch from './profile-switch.mjs'
 import sessionSwitch from './session-switch.mjs'
 import stream from './stream.mjs'
+import streamHistory from './stream-history.mjs'
 import submit from './submit.mjs'
 import transcript from './transcript.mjs'
 
 export const SCENARIOS = {
   [stream.name]: stream,
+  [streamHistory.name]: streamHistory,
   [keystroke.name]: keystroke,
   [transcript.name]: transcript,
   [multitab.name]: multitab,

apps/desktop/scripts/perf/scenarios/stream-history.mjs

@@ -0,0 +1,18 @@
+// Streaming into an ALREADY-LONG transcript. Same measurement as `stream`, but
+// the history is mounted and allowed to settle before the recorders start, so
+// what it captures is the per-delta cost that scales with transcript length —
+// the regression reported in #69120.
+//
+// Report-only (tier: manual): the number depends on how much history the host
+// can mount, so it is not gated against the committed baseline.
+
+import stream from './stream.mjs'
+
+export default {
+  name: 'stream-history',
+  tier: 'manual',
+  description: 'Streaming cost with a long settled transcript already mounted.',
+  run(cdp, opts = {}) {
+    return stream.run(cdp, { ...opts, historyTurns: Number(opts.historyTurns ?? 200) })
+  }
+}

apps/desktop/scripts/perf/scenarios/stream.mjs

@@ -70,7 +70,7 @@ const COLLECT = `
   })()
 `
 
-function analyze(data, warmupMs) {
+function analyze(data, warmupMs, extra = {}) {
   // Drop warm-up frames (recorder installs before the stream starts).
   const frames = []
   let acc = 0
@@ -102,6 +102,7 @@ function analyze(data, warmupMs) {
       intermut_p95_ms: Math.round(percentile(interMut, 0.95) * 10) / 10
     },
     detail: {
+      ...extra,
       windowS: Math.round(windowS * 10) / 10,
       avgFps: windowS ? Math.round((frames.length / windowS) * 10) / 10 : 0,
       frameHistogram: frameHistogram(frames),
@@ -128,13 +129,36 @@ export default {
     // noise unrelated to render cost).
     const chunk = opts.chunk ?? 'A streamed sentence with **bold**, `code`, and ordinary prose like a normal reply.\n\n'
     const real = Boolean(opts.real)
+    const historyTurns = Number(opts.historyTurns ?? 0)
+    const historySettleMs = Number(opts.historySettleMs ?? 1500)
 
     await cdp.send('Runtime.enable')
+
+    // Mount the settled history BEFORE the recorders start, so the measurement
+    // window contains only streaming work — not the one-off mount cost.
+    if (historyTurns > 0) {
+      if (real) {
+        throw new Error('--historyTurns is only supported by the synthetic stream path')
+      }
+
+      await cdp.eval(`window.__PERF_DRIVE__.loadTranscript(${historyTurns})`)
+      await sleep(historySettleMs)
+
+      const mounted = Number(await cdp.eval('window.__PERF_DRIVE__.snapshotMsgs()'))
+      const expected = historyTurns * 2
+
+      if (mounted !== expected) {
+        throw new Error(`expected ${expected} preloaded history messages, got ${mounted}`)
+      }
+    }
+
     await cdp.eval(RECORDERS)
 
     if (real) {
       // Backend path: fire a real prompt and wait for the stream to appear.
-      const baseCount = await cdp.eval(`document.querySelectorAll(${JSON.stringify(SELECTORS.assistantMessage)}).length`)
+      const baseCount = await cdp.eval(
+        `document.querySelectorAll(${JSON.stringify(SELECTORS.assistantMessage)}).length`
+      )
       await typeIntoComposer(cdp, opts.prompt ?? 'count from 1 to 80, one number per line', { cps: 40 })
       await cdp.eval(`(() => {
         const el = document.querySelector(${JSON.stringify(SELECTORS.composer)})
@@ -186,6 +210,6 @@ export default {
       await cdp.eval('window.__PERF_DRIVE__.reset()')
     }
 
-    return analyze(data, real ? 0 : 500)
+    return analyze(data, real ? 0 : 500, { historyTurns })
   }
 }