[imakris]

Merged: benchmark foundation and workload-v3 correction. Final continuation evidence: PR #31.

Why

The existing offscreen benchmark could miss settled layout reuse and rescan immutable fixture data, conflating fixture behavior with production work. Its frame-only mode also retained internal profiler bookkeeping inside the frame measurement. This PR establishes repeatable workloads and corrects the measurement boundary so later optimizations can be assessed against explicit controls.

Changes

• Add a 14-case catalog covering settled LINE/AREA/DOTS and stacked AREA, dense data, pan/zoom, append/ring wrap, text-disabled controls and conservative data access.
• Match the shipping widget's measured-width feedback and distinguish immutable stable-access fixtures from conservative controls.
• Drive input changes by measured-frame index; validate frame counts, publications, occupancy, overwrites, backend identity and expected scan/upload reuse.
• Run fresh paired processes in alternating A/B and B/A order, preserving commands, raw reports, per-process variation and build provenance.
• Correct frame-only timing: disable remaining renderer profiler work, stop timing/allocation boundaries before storing observations, and omit unavailable cold preparation timing.
• Advance the workload contract to v3 and reject all internal renderer scopes/observations and internal planning/preparation measurements in frame-only mode. Historical v2 evidence remains labeled v2.
• Retain focused tests, reproduction instructions, the original report and an initial continuation checkpoint. The completed 144-process continuation report is recorded in follow-up PR #31.

Production files under src/ and include/ are unchanged from the base. The production label-membership experiment and temporary same-executable diagnostic remain on separate branches.

Evidence and decision

The text workloads repeatedly save 38 allocations and 896 requested bytes per frame. In the same-executable comparison, settled text changes from 190 to 152 allocations/frame; conservative text changes from 206 to 168. No-text allocation counts remain unchanged.

The retained two-binary conservative no-text comparison is slower in all eight balanced pairs, with a +5.45% median paired change. In the corrected same-executable diagnostic, the corresponding control is evenly split, with a −0.71% median. Settled-text mean time favors the candidate by −4.24% median, but variation overlaps the identical-mode control and the paired p95 median is +0.47%.

This confirms the allocation reduction without establishing a general timing benefit. Code placement remains a hypothesis for the retained-binary penalty: the same-executable experiment also changes the harness from v2 to v3, so the cause is not isolated. The production optimization remains outside this PR.

• Completed continuation and reproduction commands
• Machine-readable measurements
• Original report

Validation and limits

• Rebuilt the corrected foundation and diagnostic using GCC 13.3.0 / Qt 6.12.0, Release, with identical explicit dependency inputs.
• 10/10 selected foundation CTest registrations passed, including the runner's 10 Python tests.
• 3/3 focused diagnostic CTest registrations passed in each mode.
• All 144 additional measured processes completed validation. Independent review recomputed the paired results and checked every raw report, actual mode marker, frame count, allocation bounds and warmup phases.
• Project/dependency source inventories and executable hashes stayed unchanged through build, tests and measurements. Published Git trees match the local trees exactly.

Measurements use explicit QRhi Null. They cover CPU-preparation-path wall times and observed render-thread C++ allocation traffic. Hardware GPU throughput, application FPS, presentation latency and pixel equivalence remain unverified.

A bounded native display attempt failed because this environment cannot create the required AF_UNIX sockets. FontRendererDrawRanges and SampleUploadReuse remain unrun here; Null is not substituted for those raster gates. Hosted native CI was not dispatched. This PR merged on 4 November 2026 at 01:53 UTC. The final report was completed after the merge and is proposed separately in PR #31.

Source identity and retained artifacts

Publication used the connected GitHub API because shell Git push credentials were unavailable. API commit identifiers differ from local identifiers; corresponding Git tree IDs were verified equal.

Purpose	Published commit	Local source revision
Original measured baseline	1ecbeab5f1a4f876e9a30a8e9697990b226d2e3a	88d459dbbad9cff1a292b3ffa87be97970d371e6
Initial recovered foundation	a15293f39b52bbee54e129d59e6739e74a2970e6	15eb736a897d147afa799470318312d913f06f13
Separate original production experiment and final original report	71df3e83b56d80b856c2bad73faa231c1200fff0	e9cf091df613709fc8013def2b47062cfbf97c06
Corrected foundation built and tested in this continuation	ca15d8cea331fb70a923bcb1a8c75ede3493882f	043e0a61d6e0a4127399aa341c397249634e4791
Same-executable diagnostic built and measured	94a3d43388f91ea5213c2057c3f5f20fad2091a2	41a0088fe1faf4a85b21638631e32b80501c6c70
Completed continuation report and measurements, in PR #31	0ba0a410e07efa11ec5620496c808f2bab02d2fc	7600d01848f56da988246562bd91e307954b5255

Diagnostic reproduction branch: diagnostic/label-membership-same-image-20261104. It is not a proposed production configuration.

The original vnm_plot-benchmarks-2026-11-04.zip preserves original Git objects, patches, all 284 original measured processes and logs. The new vnm_plot-benchmark-followup-2026-11-04.zip preserves the 144 continuation processes, fixed plans, build/test logs, assembly evidence, independent review and checksums. Source repositories, SDKs and duplicated binaries are excluded from the new raw-evidence archive.