perf(ci): startup / first-frame metric for /perf (PR6)

[azchohfi]

What

Adds a startup / first-frame section to the /perf comment — the from-scratch mount cost that every steady-state table excludes by construction (they window-average over benchmark ticks, after the mount). This is exactly where a #696-class mount/reconcile regression lives, and nothing in the harness measured it. Closes the user's originating "first frame rendered" question.

Class-B harness PR — 0 src/Reactor changes (everything under tests/stress_perf/**).

The four anchors (captured once per process, all "lower is better", ms from managed entry)

Field	What it is	Why
firstReconcileDurationMs	Reactor-isolated first-render reconcile-phase duration (the first OnRenderComplete's reconcile arg)	The diagnostic. Same phase steady-state Avg Diff averages, so it's directly comparable — but undiluted by AOT/window/XAML bootstrap, so a 2× mount regression shows at full size instead of as a single-digit-% bootstrap blip. This is the field that makes the metric catch the user's class of bug.
entryToFirstFrameMs	managed entry → first composed frame	The human-facing "first frame rendered" headline.
entryToFirstReconcileMs	managed entry → first reconcile	Guaranteed-monotonic companion.
windowOpenToFirstReconcileMs	window Activated → first reconcile	n/a-guarded secondary — Activated-vs-mount ordering is non-deterministic across launches, so it emits JSON null → renders n/a rather than a garbage negative that would poison the paired CI.

How it's wired (0 src/Reactor)

• StressPerf.Shared/StartupTiming.cs (new) — process-static monotonic entry/window-open anchors (Stopwatch.GetTimestamp()), idempotent first-wins, AOT/trim-safe.
• StressPerf.Shared/PerfTracker.cs — RecordFirstRenderIfUnset (one-shot, called before the benchmark gate so it sees the full mount); first-frame capture gated so T_firstFrame ≥ T_firstReconcile; 4 nullable accessors; JSON emission (null for an un-captured anchor).
• StressPerf.ReactorOptimized/Program.cs — MarkEntry() as the first statement of Main; subscribe PrimaryWindow.Activated → MarkWindowOpen(); RecordFirstRenderIfUnset(...) as the first line of OnRenderComplete.
• ci/PerfLib.ps1 — parse the 4 optional fields (a head that predates them → n/a, exactly like the alloc fields), aggregate them as headline keys (reuses the paired-95%-CI machinery), Format-PerfStartupSection, a dormant flag + provisional band, and an info glyph.

The section piggybacks the existing headline per-rep launches (one sample/process) — so it costs zero extra CI time, and the cold first launch is warmup-dropped automatically because it rides the same per-rep metrics object that Run-PerfBenchmark.ps1's interleave loop already discards. Run-PerfBenchmark.ps1 needs no change.

⚠️ Flags for the reviewer / coordinator

1. This introducing run renders PR-side shape only — by design. perf-compare.yml builds the baseline exe from a separate main worktree and overlays the PR via ProjectReference (not Program.cs/PerfTracker.cs). So on this PR's /perf run the main side predates the startup fields → emits null → no paired Δ (the section renders a visible n/a note: "populates on the next /perf after this lands"). The paired Δ + the identical-binary band calibration only populate on the first run after merge. The gate-clean run here can prove PR-side shape (fields present, monotonic, firstReconcile >> AvgDiff), not the paired delta.
2. firstReconcileDurationMs is directly comparable to steady-state Avg Diff (same reconcile/diff-patch phase). Expected shape: firstReconcileDurationMs >> AvgDiffMs (mount creates every control; steady ticks patch a few).
3. Informational-first (dormant flag). $StartupAutoFlag ships $false — Δ + CI are shown but no row auto-flags better-or-worse until a real-CI identical-binary ~0-false-flag band calibration (same discipline as the micro ns flag; measurement-only, never changes what merges). Startup is the noisiest axis (one sample/launch + bootstrap process-to-process variance), so provisional band ≥ the micro band.
4. Carry-for-user (please confirm): the headline entryToFirstFrameMs (managed entry → first composed frame) should match whatever the user's dashboard labels "first frame rendered". Flagging so the harness number is directly comparable to what they originally raised.

Tests

+39 PerfLib.Tests assertions: parse (incl. JSON-null window-open + pre-metric head), aggregate, render (informational/dormant + armed verdict), §11 lineage both directions, monotonic shape, cold-first-launch warmup-drop, comment ordering/footnote, info glyph. PerfLib.Tests 322 + RunPerfBenchmark.Tests 76 green.

Docs: METHODOLOGY.md + ci/README.md startup sections.

---

🤖 Draft — held for coordinator independent-verify + merge (Class-B coordinator-merge discipline).

[Copilot]

Pull request overview

Adds “startup / first-frame” telemetry to the stress-perf /perf comment pipeline so CI can see from-scratch mount cost (first reconcile and first composed frame) that steady-state windows intentionally exclude.

Changes:

• Introduces process-wide startup anchors (MarkEntry, MarkWindowOpen) and threads them through StressPerf.ReactorOptimized to capture first-reconcile + first-frame metrics.
• Extends PerfTracker to record/emit 4 optional startup metrics as JSON (null when unavailable) and updates PerfLib parsing/aggregation/rendering to show a new “Startup / first frame” table.
• Updates methodology/CI docs and expands PerfLib PowerShell tests to cover parsing, lineage/back-compat, warmup-drop behavior, and rendering.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/stress_perf/StressPerf.Shared/StartupTiming.cs	New process-static timing anchors for entry/window-open.
tests/stress_perf/StressPerf.Shared/PerfTracker.cs	One-shot capture + JSON emission of first-reconcile/first-frame metrics.
tests/stress_perf/StressPerf.ReactorOptimized/Program.cs	Wires entry/window-open and first-render capture into the harness.
tests/stress_perf/METHODOLOGY.md	Documents what the startup metrics mean and how they’re captured/consumed.
tests/stress_perf/ci/README.md	Documents the new /perf startup section and its interpretation.
tests/stress_perf/ci/PerfLib.Tests.ps1	Adds PS tests covering parsing/aggregation/rendering/lineage behaviors.
tests/stress_perf/ci/PerfLib.ps1	Parses/aggregates startup keys and renders the new table with dormant flagging.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[azchohfi]

Closing per repo-owner direction: the first-frame / startup-metric investigation has been cancelled — we are no longer chasing the first-frame bump. Effort is redirected to validating /perf's existing steady-state metrics across the perf-PR fleet so the real perf PRs can be measured meaningfully. Branch is left intact in case first-frame is ever revived.