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>

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

@@ -1,92 +0,0 @@
-import { DatabaseSync } from 'node:sqlite'
-
-import { describe, expect, it } from 'vitest'
-
-import type { GiteaIssue } from '../gitea/types.js'
-import { extractLabelFacts } from '../labels/label-schema.js'
-import { type CacheDriver, initCache, readIssue, upsertIssue } from './cache-v0.js'
-
-/** Adapt node:sqlite's DatabaseSync to the CacheDriver seam (main uses better-sqlite3). */
-function memoryDriver(): CacheDriver {
-  const db = new DatabaseSync(':memory:')
-  return {
-    exec: (sql) => db.exec(sql),
-    run: (sql, params = []) => {
-      db.prepare(sql).run(...(params as never[]))
-    },
-    get: (sql, params = []) => db.prepare(sql).get(...(params as never[])) as Record<string, unknown> | undefined,
-    all: (sql, params = []) => db.prepare(sql).all(...(params as never[])) as Record<string, unknown>[],
-  }
-}
-
-function issue(over: Partial<GiteaIssue> = {}): GiteaIssue {
-  const labels = over.labels ?? ['est/5d', 'p/1', 'deadline/hard']
-  return {
-    number: 42,
-    title: 'Monte Carlo engine',
-    body: 'percentile bands',
-    state: 'open',
-    labels,
-    facts: extractLabelFacts(labels),
-    milestone: { id: 7, title: 'P2 — Scheduler', dueOn: '2026-09-01T00:00:00Z' },
-    assignee: 'christian',
-    assignees: ['christian'],
-    createdAt: '2026-07-08T00:00:00Z',
-    updatedAt: '2026-07-08T01:00:00Z',
-    closedAt: null,
-    url: 'https://gitea/christian/commitea/issues/42',
-    ...over,
-  }
-}
-
-describe('cache-v0', () => {
-  it('mirrors one issue and reads its facts back through extractLabelFacts (acceptance)', () => {
-    const d = memoryDriver()
-    initCache(d)
-    upsertIssue(d, issue({ labels: ['est/5d', 'p/1', 'deadline/hard'] }))
-
-    const back = readIssue(d, 42)!
-    expect(back.labels).toEqual(['est/5d', 'p/1', 'deadline/hard'])
-    // facts are re-derived on read, not stored
-    expect(back.facts.estimateDays).toBe(5)
-    expect(back.facts.priority).toBe(1)
-    expect(back.facts.hardDeadline).toBe(true)
-    // the rest of the domain shape round-trips
-    expect(back.milestone).toEqual({ id: 7, title: 'P2 — Scheduler', dueOn: '2026-09-01T00:00:00Z' })
-    expect(back.assignee).toBe('christian')
-    expect(back.state).toBe('open')
-  })
-
-  it('re-derives facts from the current labels after a re-reconcile (upsert in place, no dup)', () => {
-    const d = memoryDriver()
-    initCache(d)
-    upsertIssue(d, issue({ labels: ['est/2d', 'p/3'] }))
-    // reconcile again with changed labels + closed
-    upsertIssue(d, issue({ labels: ['est/8d', 'p/1'], state: 'closed', closedAt: '2026-07-09T00:00:00Z' }))
-
-    expect(d.all('SELECT number FROM issues')).toHaveLength(1) // upsert by number, not a second row
-    const back = readIssue(d, 42)!
-    expect(back.facts.estimateDays).toBe(8)
-    expect(back.facts.priority).toBe(1)
-    expect(back.facts.hardDeadline).toBe(false) // deadline/hard dropped
-    expect(back.state).toBe('closed')
-    expect(back.closedAt).toBe('2026-07-09T00:00:00Z')
-  })
-
-  it('reads an issue with no milestone / empty labels', () => {
-    const d = memoryDriver()
-    initCache(d)
-    upsertIssue(d, issue({ number: 9, labels: [], milestone: null, assignee: null, assignees: [] }))
-    const back = readIssue(d, 9)!
-    expect(back.milestone).toBeNull()
-    expect(back.labels).toEqual([])
-    expect(back.facts.estimateDays).toBeNull()
-    expect(back.assignee).toBeNull()
-  })
-
-  it('returns null for an uncached issue', () => {
-    const d = memoryDriver()
-    initCache(d)
-    expect(readIssue(d, 999)).toBeNull()
-  })
-})

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

@@ -1,159 +0,0 @@
-/**
- * SQLite cache, v0 (#3) — a rebuildable local mirror of the reconciled backlog.
- * It is an index over the durable truth in gitea, never the source of truth (D4):
- * delete it, resync, lose nothing. This module owns the schema + the pure
- * row<->domain mappers; the actual SQLite handle is injected as a `CacheDriver`,
- * so core stays free of any native driver (better-sqlite3 lives in main; tests
- * use node:sqlite). Facts are never stored — they are re-derived from the label
- * set on read via `extractLabelFacts`, so the mirror can't drift from the label
- * semantics.
- */
-
-import type { GiteaIssue, GiteaMilestoneRef } from '../gitea/types.js'
-import { extractLabelFacts } from '../labels/label-schema.js'
-
-/**
- * The injected IO boundary: a thin synchronous SQL executor. Core writes the SQL;
- * the host binds a real driver (better-sqlite3 in the desktop main process,
- * node:sqlite in tests). Kept minimal on purpose — no ORM, no query builder.
- */
-export interface CacheDriver {
-  /** Run one or more DDL/utility statements (no params, no result). */
-  exec(sql: string): void
-  /** Execute a single parameterized write. */
-  run(sql: string, params?: readonly unknown[]): void
-  /** First row of a parameterized query, or undefined. */
-  get(sql: string, params?: readonly unknown[]): Record<string, unknown> | undefined
-  /** All rows of a parameterized query. */
-  all(sql: string, params?: readonly unknown[]): Record<string, unknown>[]
-}
-
-/** The cache schema — five tables mirroring gitea's shape. Regenerable; drop and rebuild freely. */
-export const CACHE_SCHEMA = `
-CREATE TABLE IF NOT EXISTS milestones (
-  id       INTEGER PRIMARY KEY,
-  title    TEXT NOT NULL,
-  state    TEXT,
-  due_on   TEXT
-);
-CREATE TABLE IF NOT EXISTS issues (
-  number       INTEGER PRIMARY KEY,
-  title        TEXT NOT NULL,
-  body         TEXT NOT NULL DEFAULT '',
-  state        TEXT NOT NULL,
-  labels       TEXT NOT NULL DEFAULT '[]',  -- JSON array of label names; facts re-derived on read
-  milestone_id INTEGER,
-  assignee     TEXT,
-  assignees    TEXT NOT NULL DEFAULT '[]',  -- JSON array of logins
-  created_at   TEXT,
-  updated_at   TEXT,
-  closed_at    TEXT,
-  url          TEXT,
-  FOREIGN KEY (milestone_id) REFERENCES milestones(id)
-);
-CREATE TABLE IF NOT EXISTS labels (
-  id    INTEGER PRIMARY KEY,
-  name  TEXT NOT NULL
-);
-CREATE TABLE IF NOT EXISTS comments (
-  id            INTEGER PRIMARY KEY,
-  issue_number  INTEGER NOT NULL,
-  author        TEXT,
-  body          TEXT NOT NULL DEFAULT '',
-  created_at    TEXT
-);
-CREATE TABLE IF NOT EXISTS issue_events (
-  id            INTEGER PRIMARY KEY AUTOINCREMENT,
-  issue_number  INTEGER NOT NULL,
-  type          TEXT NOT NULL,
-  at            TEXT NOT NULL
-);
-CREATE INDEX IF NOT EXISTS idx_issue_events_number ON issue_events(issue_number);
-CREATE INDEX IF NOT EXISTS idx_comments_number ON comments(issue_number);
-`
-
-/** Create the schema if absent. Idempotent. */
-export function initCache(driver: CacheDriver): void {
-  driver.exec(CACHE_SCHEMA)
-}
-
-const UPSERT_ISSUE = `
-INSERT INTO issues (number, title, body, state, labels, milestone_id, assignee, assignees, created_at, updated_at, closed_at, url)
-VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
-ON CONFLICT(number) DO UPDATE SET
-  title = excluded.title, body = excluded.body, state = excluded.state, labels = excluded.labels,
-  milestone_id = excluded.milestone_id, assignee = excluded.assignee, assignees = excluded.assignees,
-  created_at = excluded.created_at, updated_at = excluded.updated_at, closed_at = excluded.closed_at, url = excluded.url
-`
-
-const UPSERT_MILESTONE = `
-INSERT INTO milestones (id, title, state, due_on) VALUES (?, ?, ?, ?)
-ON CONFLICT(id) DO UPDATE SET title = excluded.title, state = excluded.state, due_on = excluded.due_on
-`
-
-/**
- * Mirror one reconciled issue into the cache (and its milestone, if any). Upsert
- * by `number`, so re-reconciling the same issue updates in place — never duplicates.
- */
-export function upsertIssue(driver: CacheDriver, issue: GiteaIssue): void {
-  if (issue.milestone) {
-    driver.run(UPSERT_MILESTONE, [issue.milestone.id, issue.milestone.title, null, issue.milestone.dueOn])
-  }
-  driver.run(UPSERT_ISSUE, [
-    issue.number,
-    issue.title,
-    issue.body,
-    issue.state,
-    JSON.stringify(issue.labels),
-    issue.milestone?.id ?? null,
-    issue.assignee,
-    JSON.stringify(issue.assignees),
-    issue.createdAt,
-    issue.updatedAt,
-    issue.closedAt,
-    issue.url,
-  ])
-}
-
-const READ_ISSUE = `
-SELECT i.*, m.title AS m_title, m.due_on AS m_due
-FROM issues i LEFT JOIN milestones m ON m.id = i.milestone_id
-WHERE i.number = ?
-`
-
-function str(v: unknown): string {
-  return typeof v === 'string' ? v : ''
-}
-
-/**
- * Read one mirrored issue back as a domain object, re-deriving `facts` from the
- * stored label set (so the mirror can't disagree with the label semantics).
- * Returns null when the issue isn't cached.
- */
-export function readIssue(driver: CacheDriver, number: number): GiteaIssue | null {
-  const row = driver.get(READ_ISSUE, [number])
-  if (!row) return null
-
-  const labels = (JSON.parse(str(row.labels) || '[]') as string[]) ?? []
-  const assignees = (JSON.parse(str(row.assignees) || '[]') as string[]) ?? []
-  const milestone: GiteaMilestoneRef | null =
-    row.milestone_id != null
-      ? { id: Number(row.milestone_id), title: str(row.m_title), dueOn: (row.m_due as string | null) ?? null }
-      : null
-
-  return {
-    number: Number(row.number),
-    title: str(row.title),
-    body: str(row.body),
-    state: row.state === 'closed' ? 'closed' : 'open',
-    labels,
-    facts: extractLabelFacts(labels),
-    milestone,
-    assignee: (row.assignee as string | null) ?? null,
-    assignees,
-    createdAt: str(row.created_at),
-    updatedAt: str(row.updated_at),
-    closedAt: (row.closed_at as string | null) ?? null,
-    url: str(row.url),
-  }
-}

packages/core/src/index.ts

@@ -23,8 +23,6 @@ export type {
   GiteaRequestInit,
 } from './gitea/types.js'
 
-export { CACHE_SCHEMA, initCache, readIssue, upsertIssue } from './cache/cache-v0.js'
-export type { CacheDriver } from './cache/cache-v0.js'
 export { describeChange, isLabelChange, planIssueChange, proposalsFor, summarizeChange } from './changes/apply-changes-v0.js'
 export type {
   ChangeProposal,

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
+  })
+})