Fix CI Performance Tests

[drasmart]

Changes

• Move run-performance-tests.ps1 (already coupled) from common-ci/rust to here.
• Download Asn file from Azure.
• Add dryrun and common-ci-ref to nightly-performance.yml.
• Stop hiding failures of compare-performance.ps1.

[drasmart]

Nightly Performance job:

• ✅ https://github.com/51Degrees/rust/actions/runs/34466577685

[oleksandrlazarenko-pi]

Review notes

Reviewed against main, cross-checked with common-ci/steps/compare-performance.ps1, the ip-intelligence-data submodule pins, and this workflow's Actions run history.

Blocking

1. Dropping continue-on-error: true silently disables the history upload — .github/workflows/nightly-performance.yml:162

When the compare step fails, Escape Branch Name and Upload Graphed Performance Results are skipped, because neither has if: always(). That artifact is the only source of future baselines, so the run's own figure never enters history.

Confirmed empirically on run 34464099806: steps 14 and 15 report skipped.

The consequence only bites once there are ≥10 data points: a genuine regression exits 1, the regressed figure never enters history, and every subsequent nightly then compares against the same stale baseline and fails forever. Adding if: always() to both steps keeps the loud failure without wedging the trend.

2. The ASN download resolves against the caller's CWD, not $RepoName — ci/run-performance-tests.ps1:13

The ip-intelligence-cxx/ip-intelligence-data path is relative to wherever the script happens to be invoked from rather than to $RepoName. The script's own documented local invocation (-RepoName "." from the repo root) now throws at Push-Location before any benchmark runs.

Non-blocking

• nightly-performance.yml:169 — dryrun only suppresses git push; the workflow still uploads publish_performance_results@<branch>, so a dry run on main injects its figure into the real performance history. That contradicts the new comment at lines 149-152.
• ci/run-performance-tests.ps1:24 — deleting the ASN file and calling get-lite-file-from-azure.ps1 only restores it when the ip-intelligence-data pin is ≥ 60ec85a. Older pins fetch only 51Degrees-LiteV41.ipi, leaving 51DEGREES_IPI_PATH dangling — and this PR adds the ipi-cxx-ref input that makes older pins reachable.
• ci/run-performance-tests.ps1:19 — the pre-delete cannot force a refresh: the Azure script gates its download on the .gz archive, not the .ipi, and is called without -Force. On a persisted workspace a stale archive is silently re-extracted; on a failed download the tracked LFS file is already gone.
• nightly-performance.yml:159 — the comment justifying continue-on-error: false cites an exit code 2 for infrastructure failure, but compare-performance.ps1 only ever exits 0 or 1, so infrastructure failures are indistinguishable from regressions.
• ci/missing-ci-scripts.md:67 — "behaviour of the nightly run is unchanged" and "the two copies are identical" are both falsified by this PR itself (ASN block, dryrun, refs, continue-on-error).
• nightly-performance.yml:197 — -ErrorAction Stop is a no-op, since the callee sets $ErrorActionPreference = "Stop" itself. Looks like a leftover from the reverted $PSNativeCommandUseErrorActionPreference attempt.
• nightly-performance.yml:98 — the benchmark corpus is no longer version-pinned (live Azure blob per run). A published data-file update can shift throughput past the 10%/2σ band and, with continue-on-error false, fail the nightly with no way to re-baseline.

Checked and sound

The git -C rust checkout -- examples/Cargo.lock fix (both lockfiles are tracked, /test-results is gitignored, the sys build scripts only read from the manifest dir); the throughput regexes against the examples' println! formats, and the "last match wins" comment matching the IPI example's single-then-multi report order; -DryRun $DryRun binding (-DryRun exists on common-ci main with a [bool] type); inputs.dryrun || 'false' for schedule, dispatch-true and dispatch-false; the empty ref: on actions/checkout falling back to the default branch; and the omitted description on the dryrun input.