[jogibear9988]

PR authors currently have to repeat the manual before/after process described in #64. This adds automated performance reports for PRs targeting master, initially in reporting-only mode.

Measurements

• A dedicated BenchmarkDotNet 0.15.8 suite measures construction, compilation, compilation plus first invocation, warmed delegate execution, and managed allocations for Standard and LightExpression.
• Four deterministic workloads cover scalar arithmetic, balanced arithmetic, calls/properties/conditions, and nested-lambda captures: 32 benchmark cases. Strict CompileFast and result/checksum checks reject fallback or incorrect behavior.
• Build the exact PR base and head using SDK 10.0.301. Use one frozen base-branch harness for both revisions, with isolated outputs and opposite process orders: base/PR, then PR/base.
• Compare complete benchmark/configuration identities and confidence intervals across both rounds. Missing, invalid or mismatched results are inconclusive. Hosted-runner timing differences do not fail PRs; explicit local blocking is available for later calibration.
• Default PR runs target .NET 10; manual runs support .NET 8, .NET 10, or both. Raw reports, environment information and source/library hashes are retained as artifacts for 14 days.

PR comments and permissions

Benchmark jobs run with read-only repository permissions on disposable GitHub-hosted runners. A separate workflow_run publisher executes only default-branch code and treats the small report artifact as untrusted data: bounded ZIP reads without extraction, strict schema/numeric validation, and no artifact-provided scripts or prose.

The publisher resolves the PR through GitHub's API, handles empty fork-run PR associations, checks the current head/workflow/attempt, and rejects stale results. New head commits receive a new report; older bot comments are marked Outdated and collapsed with a link to the new report. Same-head reruns update their report instead of creating duplicates.

Validation

• 13 Python policy/artifact tests pass, including reporting versus blocking exit codes, both-round confirmation, configuration mismatches and invalid archives.
• 7 mocked Node publisher tests pass, including fork resolution, stale-head/attempt rejection, supersession and rerun ordering.
• Both workflows pass actionlint.
• The four-run dry integration exercised all 32 cases against both staged libraries and is correctly classified as inconclusive rather than a performance pass.
• Full local normal .NET 10 integration completed: all 32 cases in all four counterbalanced runs; complete comparison, reporting-only exit 0. Both base and candidate also pass .NET 8 builds and all eight driver/behavior checks.
• The new GitHub-hosted performance workflow passed, as did the existing Linux and Windows build/test jobs.
• Downloaded the real GitHub report artifact and validated it through the trusted publisher: all 32 identities, correct PR/head binding, and a correctly rendered 32-row Markdown table.
• The first hosted run retained one +3.1% time/noise warning and a +1 B/op allocation warning, with no confirmed time or new-zero-path allocation regressions. Compiler sources are identical; these are useful calibration controls for the reporting-only rollout.

Rollout

Based directly on dadhi/master at 0f37356. For the installation PR, master has no harness yet, so a clearly labelled bootstrap uses the same candidate harness for both revisions. Later PRs use the base-branch harness. The publisher becomes active after its workflow is merged into the default branch; before then this PR can expose benchmark summaries/artifacts but cannot activate that trusted publisher.

This addresses #64's repeatable regression-comparison workflow. It does not enable required performance gates or implement the separate minimal-IL experiment discussed there. The regression README describes confidence/noise limits, the first-invocation JIT boundary, local replay, and the path to calibrated enforcement.

Generated comment preview from the first successful run

net10.0: Comparison complete (reporting only)

Base 0f37356fc9ca → PR ff47206815b2; SDK 10.0.301; harness ff47206815b2 (bootstrap).

Bootstrap: the base does not contain this harness yet. One candidate harness was used for both revisions; later PRs use the base-branch harness.

Backend / phase / workload	Base ns	PR ns	Time change	B/op: base → PR	Assessment
Light / Compile / Arithmetic	5606.10	5583.30	-0.4%	928.00 → 928.00	Within variation
Light / Compile / CallConditional	3277.59	3370.33	+2.9%	1380.50 → 1380.00	Within variation
Light / Compile / Closure	6189.67	6373.72	+3.1%	3033.00 → 3034.00	Warning / noise
Light / Compile / Scalar	2259.97	2279.25	+1.0%	892.50 → 892.00	Within variation
Light / CompileAndFirstInvoke / Arithmetic	44154.59	44562.84	+0.9%	928.00 → 928.00	Within variation
Light / CompileAndFirstInvoke / CallConditional	98180.77	97954.28	-0.2%	1360.00 → 1360.00	Within variation
Light / CompileAndFirstInvoke / Closure	264538.97	265196.27	+0.2%	3185.00 → 3185.00	Within variation
Light / CompileAndFirstInvoke / Scalar	10808.85	10695.95	-1.0%	872.00 → 872.00	Within variation
Light / Construct / Arithmetic	856.69	866.32	+1.1%	1296.00 → 1296.00	Within variation
Light / Construct / CallConditional	164.36	166.19	+1.1%	360.00 → 360.00	Within variation
Light / Construct / Closure	120.89	110.73	-8.1%	144.00 → 144.00	Within variation
Light / Construct / Scalar	74.31	75.60	+1.7%	120.00 → 120.00	Within variation
Light / Invoke / Arithmetic	1.92	1.92	+0.0%	0.00 → 0.00	Within variation
Light / Invoke / CallConditional	1.91	1.91	-0.0%	0.00 → 0.00	Within variation
Light / Invoke / Closure	35.66	35.10	-1.6%	184.00 → 184.00	Within variation
Light / Invoke / Scalar	1.64	1.64	+0.1%	0.00 → 0.00	Within variation
Standard / Compile / Arithmetic	6199.58	5895.11	-4.9%	945.50 → 931.00	Within variation
Standard / Compile / CallConditional	3433.97	3429.11	-0.1%	1380.50 → 1380.50	Within variation
Standard / Compile / Closure	6623.70	6660.85	+0.6%	3034.50 → 3035.50	Allocation increase
Standard / Compile / Scalar	2478.58	2478.54	+0.0%	888.50 → 888.50	Within variation
Standard / CompileAndFirstInvoke / Arithmetic	44495.31	44667.07	+0.4%	928.00 → 928.00	Within variation
Standard / CompileAndFirstInvoke / CallConditional	99627.55	99756.15	+0.1%	1360.00 → 1360.00	Within variation
Standard / CompileAndFirstInvoke / Closure	264921.79	265842.96	+0.3%	3184.00 → 3184.00	Within variation
Standard / CompileAndFirstInvoke / Scalar	10922.82	10995.30	+0.7%	872.00 → 872.00	Within variation
Standard / Construct / Arithmetic	1321.85	1320.44	-0.1%	1808.00 → 1808.00	Within variation
Standard / Construct / CallConditional	619.67	603.20	-2.6%	680.00 → 680.00	Within variation
Standard / Construct / Closure	346.21	346.00	-0.0%	440.00 → 440.00	Within variation
Standard / Construct / Scalar	263.63	266.20	+1.0%	416.00 → 416.00	Within variation
Standard / Invoke / Arithmetic	1.92	1.92	+0.1%	0.00 → 0.00	Within variation
Standard / Invoke / CallConditional	1.91	1.92	+0.1%	0.00 → 0.00	Within variation
Standard / Invoke / Closure	35.46	35.20	-0.7%	184.00 → 184.00	Within variation
Standard / Invoke / Scalar	1.64	1.64	+0.0%	0.00 → 0.00	Within variation

Two opposite-order process rounds on one GitHub-hosted runner; BDN confidence intervals must be separated in both rounds to classify a >5% time regression. A ≥3% change without that evidence is a warning, not proof. Five measurement iterations/round; hosted-runner noise remains.

Compile excludes delegate invocation; CompileAndFirstInvoke includes the first call/JIT of a new delegate. Invoke is warmed; closure invocation includes creating and calling the returned delegate. MemoryDiagnoser reports managed B/op. Loop/checksum overhead is shared by both revisions.

Workflow run and complete raw artifacts

[dadhi]

Let merge, thank you!