[imakris]

Why

PR #30 merged the benchmark foundation and workload-v3 measurement correction while the continuation report was being completed. This follow-up records the completed investigation and preserves all outcomes needed to interpret the production label-membership experiment.

Changes

Update only benchmark/results/2026-11-04-followup/README.md and measurements.json with all 144 additional measured processes, fixed plans, paired statistics, exact source/binary identities, validation results, limits and reproduction commands.

• The retained two-binary conservative no-text penalty reproduces: +5.45% median paired change, slower in all eight balanced pairs.
• The corrected same-executable no-text control has mixed signs: −0.71% median, four slower and four faster pairs.
• The allocation improvement is exact: 38 allocations and 896 requested bytes saved per frame in both measured text workloads; no-text allocations are unchanged.
• Settled-text mean timing favors the candidate by −4.24% median, but overlaps the identical-mode control's variation. The paired p95 median is +0.47%. No general speedup or tail-latency improvement is claimed.
• Code placement remains a hypothesis; historical v2 and corrected v3 campaigns do not isolate that cause.
• The production optimization remains separate from the benchmark foundation. The same-executable switch is diagnostic-only.

Complete report and reproduction commands

Machine-readable measurements

Validation

The compiled source is already included in merged PR #30. This PR changes documentation and recorded measurements only.

• Corrected foundation: 10/10 selected CTest registrations passed, including 10 Python checks.
• Diagnostic: 3/3 focused CTest registrations passed in each mode.
• All 144 measured processes completed their validations. Independent review recomputed the paired results and checked raw reports, actual mode markers, hashes, frame counts, allocation bounds and warmup phases.
• A second review checked this report against the manifests and summaries, verified reproduction commands and confirmed no production source changes.
• The published Git tree cce224f2c34b55fd093ac16b3eddef07842f5a71 exactly matches local report revision 7600d01848f56da988246562bd91e307954b5255. API-created commit identifiers differ.

All timings use QRhi Null, so native rendering, hardware GPU performance and pixel equivalence remain unqualified. The local native display attempt was blocked by AF_UNIX socket permissions; FontRendererDrawRanges and SampleUploadReuse were not substituted with Null and remain outstanding. No hosted native CI was dispatched.

Reproduction material

• Corrected compiled foundation: ca15d8cea331fb70a923bcb1a8c75ede3493882f, merged via PR Establish repeatable plot benchmarks and retain optimization evidence #30.
• Diagnostic-only source: 94a3d43388f91ea5213c2057c3f5f20fad2091a2, branch diagnostic/label-membership-same-image-20261104.
• vnm_plot-benchmark-followup-2026-11-04.zip retains all 144 raw process results, fixed plans, build/test logs, assembly evidence, independent review and checksums. Source repositories, SDKs and duplicated binaries are excluded.
• The original vnm_plot-benchmarks-2026-11-04.zip retains the original Git objects, binaries, patches and original 284-process evidence.