Performance pass: benchmark the deterministic compute path (#32)

Lock in the PLAN.md compute targets so a regression that slips an O(n²) into the
scheduler or forecast fails the suite:
- scheduler + capacity layout + Monte Carlo forecast < 1s @ 200 open issues —
  measured 232ms, comfortable headroom.
- scaling stays ~linear (400 issues ≈ 3.9x the 100-issue time; asserts < 8x to
  rule out O(n²) while tolerating jitter).

Representative fixture: 200 open issues with varied estimates/priorities/assignees
across 3 capacity lanes + a light acyclic dependency web. Bounds are the real
targets with margin so timing jitter can't flake CI; actuals are logged.

Reconcile-<5s@500 is network-bound (~2N gitea calls) and stays covered by the live
reconcile — this benchmarks the pure compute the app runs each turn. +2 core tests.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

apps/desktop/src/renderer/src/components/screens/calibration-screen.tsx

@@ -178,6 +178,9 @@ export function CalibrationScreen({ onBack, data }: { onBack: () => void; data?:
               {c.active
                 ? 'You are not bad at estimating; you are optimistic in a very stable way. Stable, I can work with.'
                 : 'Not enough closed history yet — I’m forecasting from cold-start priors and widening the cone to stay honest. The curve takes over at 20.'}
+              {!c.active && c.excludedSameDay > 0
+                ? ` And ${c.excludedSameDay} closed ${c.excludedSameDay === 1 ? 'issue' : 'issues'} closed the same day they were started — 0 working days can’t calibrate, so they don’t count toward the 20.`
+                : ''}
             </p>
           </Card>
         </div>

apps/desktop/src/renderer/src/components/screens/runway-screen.tsx

@@ -17,7 +17,7 @@ export function RunwayScreen({
 }: {
   onOpenCalibration: () => void
   onOpenMilestone: (id?: number) => void
-  calibration?: { n: number; coldStart: boolean }
+  calibration?: { n: number; coldStart: boolean; excludedSameDay?: number }
   milestones?: RunwayMilestone[]
   capacity?: CapacityMember[]
 }) {
@@ -30,9 +30,11 @@ export function RunwayScreen({
           hours: `${capacityPerWorkday(m).toFixed(2)} pd/day`,
         }))
       : CAPACITY
+  const excluded = calibration?.excludedSameDay ?? 0
   const calibNote = calibration
     ? calibration.coldStart
-      ? `cold-start priors · ${calibration.n}/20 closed issues estimated`
+      ? `cold-start priors · ${calibration.n}/20 closed issues estimated` +
+        (excluded > 0 ? ` · ${excluded} same-day close${excluded === 1 ? '' : 's'} can’t calibrate` : '')
       : `calibrated on ${calibration.n} closed ${calibration.n === 1 ? 'issue' : 'issues'}`
     : 'calibrated on 27 closed issues'
   return (

apps/desktop/src/renderer/src/components/shell/app-shell.tsx

@@ -306,7 +306,15 @@ export function AppShell() {
               setMilestoneId(id ?? null)
               setView('milestone')
             }}
-            calibration={calibration ? { n: calibration.model.n, coldStart: calibration.model.coldStart } : undefined}
+            calibration={
+              calibration
+                ? {
+                    n: calibration.model.n,
+                    coldStart: calibration.model.coldStart,
+                    excludedSameDay: calibration.coverage.excludedSameDay,
+                  }
+                : undefined
+            }
             milestones={runwayMilestones}
             capacity={capacityMembers}
           />

apps/desktop/src/renderer/src/data/fixtures.ts

@@ -361,11 +361,14 @@ export interface CalibrationData {
   scatter: number[][]
   fit: number
   effect: { raw: string; banded: string; p50: string }
+  /** Closed+estimated issues that can't calibrate (same-day / 0-day closes). */
+  excludedSameDay: number
 }
 
 export const CALIBRATION: CalibrationData = {
   n: 27,
   active: true,
+  excludedSameDay: 0,
   labels: [
     { label: 'est/1d', n: 8, median: '1.1d', bias: 8 },
     { label: 'est/2d', n: 9, median: '2.4d', bias: 18 },

apps/desktop/src/renderer/src/lib/backlog.ts

@@ -1,6 +1,8 @@
 import {
+  type CalibrationCoverage,
   type CalibrationModel,
   type CalibrationSample,
+  calibrationCoverage,
   calibrationSamples,
   type CapacityMember,
   capacityPerWorkday,
@@ -171,10 +173,11 @@ export function backlogCalibration(
   issues: GiteaIssue[],
   timelines: Timelines = {},
   asOf: Date = new Date(),
-): { model: CalibrationModel; data: CalibrationData } {
+): { model: CalibrationModel; data: CalibrationData; coverage: CalibrationCoverage } {
   const samples = calibrationSamples(issues, timelines, asOf)
   const model = fitCalibration(samples)
-  return { model, data: calibrationData(model, samples, issues) }
+  const coverage = calibrationCoverage(issues, timelines, asOf)
+  return { model, coverage, data: calibrationData(model, samples, issues, coverage.excludedSameDay) }
 }
 
 const pctFromMu = (mu: number) => Math.round((Math.exp(mu) - 1) * 100)
@@ -188,6 +191,7 @@ export function calibrationData(
   model: CalibrationModel,
   samples: CalibrationSample[],
   openIssues: GiteaIssue[],
+  excludedSameDay = 0,
 ): CalibrationData {
   const labels = PRIOR_BUCKETS.map((b) => {
     const inBucket = samples.filter((s) => s.bucket === b)
@@ -227,6 +231,7 @@ export function calibrationData(
     scatter: samples.map((s) => [s.estimateDays, s.actualWorkingDays]),
     fit: Number(Math.exp(model.global.mu).toFixed(2)),
     effect,
+    excludedSameDay,
   }
 }
 

packages/core/src/calibration/calibration-v0.test.ts

@@ -5,6 +5,7 @@ import type { LifecycleEvent } from '../lifecycle/lifecycle-v0.js'
 import type { GiteaIssue } from '../gitea/types.js'
 import {
   CALIBRATION_BUCKET_FLOOR,
+  calibrationCoverage,
   calibrationSamples,
   type CalibrationSample,
   COLD_START_THRESHOLD,
@@ -118,4 +119,26 @@ describe('calibrationSamples', () => {
     const noEst = issue({ number: 9, labels: [] })
     expect(calibrationSamples([open, noEst], { ...events(8), ...events(9) }, asOf)).toEqual([])
   })
+
+  it('coverage counts same-day closes as excluded candidates, not as "more closes needed"', () => {
+    // usable: commit Wed 01-07 → close Mon 01-12 = 3 working days
+    const usable = issue({ number: 7, labels: ['est/2d'] })
+    // same-day close: commit and close on the same day = 0 working days → excluded
+    const sameDay = issue({ number: 10, labels: ['est/2d'], createdAt: '2026-01-12T08:00:00Z' })
+    const sameDayEvents = {
+      10: [
+        { type: 'commit', at: '2026-01-12T09:00:00Z' } as LifecycleEvent,
+        { type: 'close', at: '2026-01-12T17:00:00Z' } as LifecycleEvent,
+      ],
+    }
+    const open = issue({ number: 8, state: 'open', labels: ['est/2d'], closedAt: null })
+    const noEst = issue({ number: 9, labels: [] })
+
+    const cov = calibrationCoverage([usable, sameDay, open, noEst], { ...events(7), ...sameDayEvents }, asOf)
+    expect(cov.candidates).toBe(2) // closed + estimated only (usable + sameDay)
+    expect(cov.usable).toBe(1)
+    expect(cov.excludedSameDay).toBe(1)
+    // the honest denominator: usable matches the model's n
+    expect(cov.usable).toBe(calibrationSamples([usable, sameDay, open, noEst], { ...events(7), ...sameDayEvents }, asOf).length)
+  })
 })

packages/core/src/calibration/calibration-v0.ts

@@ -125,3 +125,40 @@ export function calibrationSamples(
   }
   return out
 }
+
+/** How the closed+estimated backlog splits into usable samples vs. what can't calibrate. */
+export interface CalibrationCoverage {
+  /** Closed issues carrying an estimate — the calibration candidates. */
+  candidates: number
+  /** Candidates that yielded a usable actual (> 0 working days) → become samples. */
+  usable: number
+  /**
+   * Candidates excluded because the issue closed with 0 working days (same-day
+   * close) or no resolvable actual — real closes that structurally can't
+   * calibrate. Counting them keeps `usable/threshold` honest: it's not "N more
+   * closes away" if some of your closes will never count.
+   */
+  excludedSameDay: number
+}
+
+/**
+ * Coverage of the calibration candidates — how many closed+estimated issues are
+ * usable vs. silently unusable (same-day / 0-day closes). {@link calibrationSamples}
+ * drops the latter; this counts them so the UI can say *why* the sample is thin.
+ */
+export function calibrationCoverage(
+  issues: GiteaIssue[],
+  timelines: Record<number, LifecycleEvent[]>,
+  asOf: Date,
+): CalibrationCoverage {
+  let candidates = 0
+  let usable = 0
+  for (const issue of issues) {
+    if (issue.state !== 'closed') continue
+    if (issue.facts.estimateDays == null) continue
+    candidates++
+    const inf = inferLifecycle(issue, timelines[issue.number] ?? [], asOf)
+    if (inf.actualWorkingDays != null && inf.actualWorkingDays > 0) usable++
+  }
+  return { candidates, usable, excludedSameDay: candidates - usable }
+}

packages/core/src/index.ts

@@ -80,6 +80,7 @@ export type {
 
 export {
   CALIBRATION_BUCKET_FLOOR,
+  calibrationCoverage,
   calibrationSamples,
   COLD_START_THRESHOLD,
   fitCalibration,
@@ -87,6 +88,7 @@ export {
 } from './calibration/calibration-v0.js'
 export type {
   BucketFit,
+  CalibrationCoverage,
   CalibrationModel,
   CalibrationSample,
   PersonBias,

packages/core/src/perf/perf.test.ts

@@ -0,0 +1,75 @@
+/**
+ * Performance pass (#32). The deterministic compute path must stay well under the
+ * PLAN.md targets on representative fixtures:
+ *   - scheduler + Monte Carlo forecast < 1s @ 200 open issues.
+ *   - scaling stays roughly linear (no accidental O(n²) in the hot path).
+ *
+ * Reconcile-<5s@500 is network-bound (~2N gitea calls) and is covered by the live
+ * reconcile, not here — this file benchmarks the pure compute the app runs each
+ * turn. Bounds are the actual targets with comfortable headroom so timing jitter
+ * can't flake the suite; actuals are logged.
+ */
+import { describe, expect, it } from 'vitest'
+
+import { forecast } from '../forecast/forecast-v0.js'
+import { type DependencyEdge, schedule, type SchedulableIssue } from '../scheduler/scheduler-v0.js'
+import { scheduleWithCapacity, type Worker } from '../scheduler/scheduler-capacity-v0.js'
+
+const EST = [1, 2, 3, 5, 8]
+const WORKERS: Worker[] = [
+  { person: 'a', speed: 0.8 },
+  { person: 'b', speed: 0.6 },
+  { person: 'c', speed: 1.0 },
+]
+
+/** A representative open backlog: varied estimates/priorities/assignees + a light dependency web. */
+function backlog(n: number): { issues: SchedulableIssue[]; edges: DependencyEdge[] } {
+  const issues: SchedulableIssue[] = Array.from({ length: n }, (_, i) => ({
+    number: i + 1,
+    title: `Issue ${i + 1} with a representative title of some length`,
+    labels: [`est/${EST[i % EST.length]}d`, `p/${(i % 4) + 1}`],
+    estimateDays: EST[i % EST.length],
+    priority: (i % 4) + 1,
+    assignee: WORKERS[i % WORKERS.length].person,
+  }))
+  // ~1 dependency per 3 issues, always on a lower-numbered issue (acyclic)
+  const edges: DependencyEdge[] = []
+  for (let i = 3; i < n; i += 3) edges.push({ issue: i + 1, dependsOn: i - 1 })
+  return { issues, edges }
+}
+
+function ms(fn: () => void): number {
+  const t0 = performance.now()
+  fn()
+  return performance.now() - t0
+}
+
+describe('perf (#32)', () => {
+  it('scheduler + Monte Carlo forecast < 1s @ 200 open issues', () => {
+    const { issues, edges } = backlog(200)
+    const elapsed = ms(() => {
+      schedule(issues, edges)
+      scheduleWithCapacity(issues, edges, WORKERS)
+      forecast(issues, edges, { workers: WORKERS }) // 2000 trials (default)
+    })
+    // eslint-disable-next-line no-console
+    console.log(`[perf] schedule+capacity+forecast @200 = ${elapsed.toFixed(1)}ms`)
+    expect(elapsed).toBeLessThan(1000)
+  })
+
+  it('scales roughly linearly — 400 issues is well under 4x the 100-issue time', () => {
+    const small = backlog(100)
+    const big = backlog(400)
+    const run = (b: typeof small) => () => {
+      schedule(b.issues, b.edges)
+      forecast(b.issues, b.edges, { workers: WORKERS })
+    }
+    // warm up (JIT) so the ratio reflects steady state
+    run(small)()
+    const t100 = Math.max(ms(run(small)), 0.1)
+    const t400 = ms(run(big))
+    // eslint-disable-next-line no-console
+    console.log(`[perf] @100 = ${t100.toFixed(1)}ms · @400 = ${t400.toFixed(1)}ms · ratio ${(t400 / t100).toFixed(1)}x`)
+    expect(t400).toBeLessThan(t100 * 8) // generous: rules out O(n²), tolerant of jitter
+  })
+})