mirror of
https://github.com/stablyai/orca.git
synced 2026-10-20 08:03:28 +00:00
test(renderer): make detected index perf gate deterministic
This commit is contained in:
@@ -129,7 +129,7 @@ describe('detected worktree index performance', () => {
|
||||
expect(walkedReads).toBeGreaterThan(100_000)
|
||||
})
|
||||
|
||||
it('keeps the lookup path an order of magnitude cheaper than the catalog walk', () => {
|
||||
it('records lookup-path timing without making wall clock a CI gate', () => {
|
||||
const walkPool = freshCatalogPool()
|
||||
const walkMs = measureSweeps((sweep) => {
|
||||
const catalog = walkPool[sweep]!
|
||||
@@ -194,9 +194,6 @@ describe('detected worktree index performance', () => {
|
||||
`${JSON.stringify(report, null, 2)}\n`
|
||||
)
|
||||
|
||||
// Only the warm ratio is asserted: a republished catalog pays one rebuild that costs about
|
||||
// what a single sweep's walk did, so the fresh leg is parity and is recorded, not gated.
|
||||
// The measured warm margin is ~100x, so this bound cannot flake under CI load.
|
||||
expect(walkWarmMs / warmMs).toBeGreaterThan(10)
|
||||
// Why: shared-runner timing is noisy; the read-count test above is the deterministic CI gate.
|
||||
})
|
||||
})
|
||||
|
||||
Reference in New Issue
Block a user