Skip to content

Beautiful, AI-native markdown IDE and LLM wiki

4.7 KB
Raw Blame
{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "schemaVersion": 1,
  "qa022": {
    "p50Ms": 17,
    "capturedAt": "2026-04-17",
    "capturedFrom": "local seed (3 consecutive runs, port 22437) — replace with median-of-5 CI runs post-merge per perf-baseline-update.md",
    "notes": "FR-21 chunked-insertion frame-time p50 during a 1MB Source paste. Local consistently lands at 16.7ms (60fps tick); CI runners are typically 1.5-2x slower (33-67ms p50 observed across 6 sampled runs 2026-05-07). The baseline-relative assertion is `p50 < max(2 * p50Ms, 80)` — 80ms is the absolute floor calibrated against CI runner variance distribution (the prior 32ms floor sat below CI's noise floor); 2x p50Ms is the regression ceiling. At the current local-seed p50Ms=17, the 80ms floor dominates the assertion; rebaselining against median-of-5 CI samples (expected ≈40ms) shifts the dominant term to 2 × baseline. Key naming: `qa022` (lowercase, no dash) vs the `QA-022` ID used in SPEC.md / test names / documentation — shell-quoting friendliness, see perf-baseline-update.md §Schema."
  },
  "v2g1WarmSwitch": {
    "p50Ms": null,
    "absoluteFloorMs": 200,
    "capturedAt": "2026-04-20",
    "capturedFrom": "V2 SPEC §7 M1 + §2 G1 target — warm-switch (cache-hit) wall-clock P95 < 200ms prod for docs ≤200 views. Baseline to be captured by median-of-5 prod CI runs post-V2-integration (US-008 + wired scenarios). Absolute floor 200ms = spec G1.",
    "notes": "V2 editor cache cache-hit path. Consumer: packages/app/tests/perf/scenarios/warm-switch-cached.ts (to be authored in US-008 integration). Assertion shape: `p50 < max(2 * p50Ms, 200)` once p50Ms is captured."
  },
  "v2g2ColdPoolWarm": {
    "p50Ms": null,
    "absoluteFloorMs": 500,
    "capturedAt": "2026-04-20",
    "capturedFrom": "V2 SPEC §7 M2 + §2 G2 target — cold-pool-warm P95 < 500ms prod for docs ≤200 views. Baseline per Part B measurements: README 192ms, STORIES 541ms (pre-V2). Target post-V2: STORIES < 300ms.",
    "notes": "Existing cold-pool-warm.ts scenario; no baseline until post-V2-integration. Absolute floor 500ms = spec G2."
  },
  "v2g3ColdLoad": {
    "p50Ms": null,
    "absoluteFloorMs": 1500,
    "capturedAt": "2026-04-20",
    "capturedFrom": "V2 SPEC §7 M3 + §2 G3 target — cold-load P95 < 1500ms prod for docs up to 500KB/200 views. Baseline pre-V2: STORIES 1845ms prod. Target post-Option-E: STORIES < 1500ms.",
    "notes": "Existing cold-load scenario + Option E fallback (US-010+US-011). Absolute floor 1500ms = spec G3."
  },
  "v2g4ModeToggle": {
    "p50Ms": null,
    "absoluteFloorMs": 200,
    "capturedAt": "2026-04-20",
    "capturedFrom": "V2 SPEC §7 M4 + §2 G4 target — mode-toggle P95 < 200ms prod. CV:hidden probe AC22 MET at 204ms first / 115ms repeat dev. Target: < 200ms prod.",
    "notes": "US-009 .ok-mode-hidden CSS swap. Absolute floor 200ms = spec G4."
  },
  "f0ShellSnapCoupling": {
    "couplingP50": 3.18,
    "couplingObservedMax": 4.08,
    "couplingCeiling": 5.5,
    "absoluteBackstopMs": 1500,
    "capturedAt": "2026-08-05",
    "capturedFrom": "local M-series, 40 warm-nav samples across two contention arms — 8 samples at workers=1 and 32 at workers=4 (matching the CI worker count). Ratio-shaped baseline, so it is NOT subject to the median-of-5-CI-runs rule that absolute p50Ms keys are; see perf-baseline-update.md §Ratio-shaped baselines.",
    "notes": "F0 shell-snap coupling = shellSnap(mark-heavy doc) / shellSnap(trivial doc), both measured in the same run on the same machine. Contention arm moved the absolute legs +30.5% (heavy) and +35.2% (light) but the quotient only -4.4% (3.374 -> 3.226), with near-identical spread (2.87-4.08 vs 2.80-4.04) — that invariance is the whole point of the ratio and is what a fixed-millisecond budget could not deliver (the prior 500ms constant was 1.9x local-healthy but ~0.99x CI-healthy, so it went red on correct code). Ceiling 5.5 is set so that a doubling of the coupling fails from ANYWHERE in the observed healthy spread, not just from the median: 2 x the lowest healthy sample (2.804) is 5.608, which still trips. That whole-spread bar is required because the suite runs retries:2 with failOnFlakyTests:false, so a gate that only reddens on some runs of a regressed system goes green on retry and detects nothing. Cost of the tighter ceiling is headroom: 1.35x above the worst observed healthy sample (vs 1.47x at ceiling 6), still ~6.8 sigma out on the measured spread. absoluteBackstopMs 1500 covers only the regression class that inflates both legs equally; it is ~3x the observed CI healthy p90 and sits below the 2s in-page poll ceiling so it reports as a budget failure rather than a lost measurement. Recalibrate from the [F0] log line, not from a single red run."
  }
}