perf(ci): wire StressPerf.KeyedList into /perf as a third macro leg

[azchohfi]

What

Wires the already-merged (#694) StressPerf.KeyedList workload into the /perf
slash-command as a third interleaved A/B macro leg, so /perf can measure the
keyed child-diff path that the StocksGrid macro workload never exercises.

This is the next step in the /perf harness fleet (after #691 alloc metric + reps/CI,
#693 micro-suite, #697 skip-floor column) — it closes the last big macro blind spot.

Why

The headline StocksGrid workload (StressPerf.ReactorOptimized) mutates cells in
place by index, so its child diff always takes
ChildReconciler.ReconcilePositional. The reconciler's keyed arm —
ReconcileKeyed → ReconcileKeyedMiddle, the LIS-based minimal-move pass — is
invisible to it by construction. That is the same class of structural blind spot
the methodology validation (n=12 local analysis) flagged on the original 2-run-median
harness: optimizations the macro workload can't route through simply don't appear,
so a "within noise" verdict can't be trusted to mean "no win."

StressPerf.KeyedList renders ~500 stably keyed rows and reorders / inserts /
removes them each tick (deterministic: fixed RNG seed, constant N via paired
insert+remove, content-stable labels) — driving the keyed LIS diff on every tick.
That makes it the sensitive macro signal for the in-flight keyed work the positional
cells can never reach:

• P2 / perf: eliminate per-diff allocations in ChildReconciler + KeyedListDiff #657 (keyed-list diff)
• keyed structural-skip P4-C2

How

Measurement tooling only — 0 src/Reactor changes (7 files, all under
tests/stress_perf/** + .github/workflows/perf-compare.yml):

• Run-PerfBenchmark.ps1 — AppRegistry KeyedList entry; -IncludeKeyedList
  (default on, so fleet PRs' /perf auto-measures the keyed path) + 'KeyedList'
  in the local -Apps set; best-effort build (try/catch → omit the table on
  failure, exactly like the micro leg, so a keyed build failure never breaks the
  StocksGrid comparison); a third interleaved leg at the headline -Percent with
  the same pair-aligned drop-both collection and the full Warmup+Reps budget; ctx
  samples + result.json + Format-PerfComment threading.
• PerfLib.ps1 — Format-PerfKeyedListSection renders the four headline metrics
  with the same paired-Δ 95% CI / direction-aware status as the headline table
  (empty array when either side is null → caller renders nothing); Format-PerfComment
  gains -MainKeyed/-PrKeyed, rendered after the Allocation table, before the
  micro-suite.
• Tests — PerfLib.Tests.ps1 (216 assertions, +13) covers the new section,
  direction-awareness, -Percent threading, and comment placement;
  RunPerfBenchmark.Tests.ps1 (34 assertions, +7) covers the static wiring
  contract (param default, registry mapping, interleaved leg, comment threading).
• Docs — ci/README.md params table + "The comment" bullet; METHODOLOGY.md
  keyed-list subsection; perf-compare.yml header + timeout-minutes 60 → 75
  (third macro leg adds a build + ~Reps runs alongside the Rust cold-build).

Verification

• Both perf-lib suites green: PerfLib.Tests.ps1 216/216, RunPerfBenchmark.Tests.ps1
  34/34.
• Local Run-PerfBenchmark.ps1 compare-mode smoke (stub harness exes, since the
  real WinUI exe can't build on this ARM64/OneDrive box): the keyed-list table renders
  in the correct position with a direction-correct paired Δ + 95% CI, and is omitted
  under -IncludeKeyedList:$false (skip-floor + the rest unaffected).

Scope / discipline

Harness change — no /perf run needed on this PR (it changes /perf itself; we re-run
/perf on the fleet after it lands). DRAFT until the gate is clean. I own the
orchestration files (perf-compare.yml, Run-PerfBenchmark.ps1, PerfLib.ps1) to keep
them conflict-free across the roadmap.

[Copilot]

Pull request overview

Wires the StressPerf.KeyedList macro workload into the /perf CI orchestration as a third interleaved A/B leg, so perf-compare can measure the reconciler’s keyed child-diff path (which the positional StocksGrid workload cannot exercise).

Changes:

• Add -IncludeKeyedList (default $true) and registry wiring for StressPerf.KeyedList, with best-effort build and an interleaved main/pr run loop that preserves paired-sample alignment.
• Add PerfLib rendering for a new “Keyed-list workload” section (same paired-Δ 95% CI + direction-aware status as the headline table) and thread it into the sticky comment layout.
• Update tests/docs and increase the perf workflow timeout to account for the added build + macro leg.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/stress_perf/METHODOLOGY.md	Documents why a keyed-list macro leg is needed and how it’s measured/controlled.
tests/stress_perf/ci/Run-PerfBenchmark.ps1	Adds KeyedList registry entry, -IncludeKeyedList, best-effort build, and third interleaved A/B leg; threads aggregates into comment + result.json.
tests/stress_perf/ci/PerfLib.ps1	Adds Format-PerfKeyedListSection and renders it in Format-PerfComment (after alloc, before micro-suite).
tests/stress_perf/ci/PerfLib.Tests.ps1	Adds coverage for keyed-list section rendering/omission, direction-awareness, and placement in the comment.
tests/stress_perf/ci/RunPerfBenchmark.Tests.ps1	Adds AST/string-contract tests for the new param defaults, registry mapping, interleaved tags, and comment threading.
tests/stress_perf/ci/README.md	Updates CLI parameter docs and comment layout explanation to include the keyed-list table.
.github/workflows/perf-compare.yml	Updates header docs and increases timeout to 75 minutes to cover the added macro leg/build.

---

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