perf: reuse exact-input tablet pointer geometry

[yeahjack]

Cache exact-input Mver tablet geometry

Independent branch: perf/mver-geometry-cache-20261005, based on
666999650f8afb405fa34dbf1ee0b98cd7145117.

Scope and preservation

• Cache the last successful tablet geometry using exact representations of all
  six float inputs and handedness; check independently before and after keys
• Keep the original mouse geometry computation path and original mouse after-key
  no-op. An all-mode prototype was narrowed after measurable mouse-miss overhead
• Preserve all geometry arithmetic, textures, device/button selection, color,
  scale, vertex generation, GL state changes and ordered draw calls
• Do not cache failures; invalidate on clear/load; detect changes between phases
• Private overlay size on this Linux x86-64 build: 360 → 600 bytes, +240 bytes
  per instance in its existing allocation; no new heap allocation calls

Verification

• Full diagnostic Release build, GCC 14.2.0, warnings as errors: pass
• All 7 CTests: pass; source-line, localization, runtime safety policies: pass
• 41,066 byte-exact geometry comparisons, 20,618 counted computations: zero failures
• 8,308 draw-phase comparisons, 44,765,568 exact GL trace/payload bytes: zero differences
• Frozen renderer body verified byte-for-byte against upstream, apart from test
  preamble symbol remapping
• Dense input grids, handedness, all buttons, mouse/tablet, missing textures,
  changing color/scale/reference size/offsets, signed zero, subnormals, infinities,
  distinct NaN payloads, deterministic arbitrary float patterns, failed geometry,
  load/reload/clear and valid-cache-to-failed-load transitions covered
• Focused ASan/UBSan checks of changed code and geometry calculation pass.
  Other core/archive dependencies remain Release-built. LeakSanitizer cannot run
  under this sandbox's ptrace setup; leak detection is not claimed
• Independent read-only review found no production equivalence issue

Final measurements

Intel Xeon Platinum 8573C, Linux x86-64, GCC 14.2.0, Release -Os + IPO for
baseline/candidate/core; deliberately non-IPO opaque GL/full-geometry sinks.
Pinned to allowed CPU 0. Three runs, each with 3 warm-up and 15 measured rounds.
Each variant performs 2,000 operations per trial in alternating 100-operation
AB/BA blocks. Project builds/tests paused during measurements. Shared-host noise
remains; all final raw data and analysis code are committed.

CPU timings, microseconds per geometry call or complete two-phase CPU draw:

Scenario	Upstream median	Candidate median	Upstream p95 batch mean	Candidate p95 batch mean	Paired median delta	Paired change
Tablet, stationary	46.390	6.722	51.682	7.992	-39.247	-85.10%
Tablet, changing each frame	45.592	25.781	50.259	31.347	-19.138	-42.02%
Tablet, changed again between phases	45.787	45.558	54.095	51.564	-0.029	-0.07%
Mouse, stationary control	27.226	26.533	31.635	31.225	+0.090	+0.37%
Mouse, changing control	26.232	25.948	29.826	29.954	+0.146	+0.52%
Geometry helper, stationary	18.948	0.138	22.624	0.249	-18.711	-99.26%
Geometry helper, changing	19.354	19.233	23.003	23.536	-0.180	-0.95%

Wall-clock medians (upstream → candidate, µs): stationary tablet 46.378 → 6.727;
changing tablet 45.578 → 25.779; mid-frame-changing tablet 45.792 → 45.544;
stationary mouse 27.235 → 26.587; changing mouse 26.233 → 26.001;
stationary geometry 18.939 → 0.135; changing geometry 19.345 → 19.224.
Full wall/CPU p95 and paired statistics are in the committed summary CSV.

Conservative interpretation: ~85% less CPU preparation for repeated tablet
inputs and ~42% less when tablet coordinates change once per frame. No claimed
benefit for mouse controls or two-miss tablet frames; their differences are
noise-scale. The original all-mode prototype produced +1.17%, +3.31%, +2.56%
paired moving-mouse CPU overhead in the improved protocol, so that path was
excluded. Exploratory runs 1–5 remain in the shared study results, not this PR.

These are component costs, not end-to-end input latency. p95 is the percentile
of batch averages, not input-event tail latency. No actual native window, GPU
rendering, Cubism SDK, OS input-capture, compositor or display latency was tested.
Identical ordered GL commands/payloads support unchanged visual output under
identical GPU state; actual native-window screenshots are not claimed.

Reproduction

See tests/performance/MVER.md and
tests/performance/results/mver-20261005/README.md.
Final raw files: run-6.csv, run-7.csv, run-8.csv; aggregate:
summary-final.csv. Local study also retains all eight raw runs and build/test
logs under the shared results directory.

Re-review status (2026-10-05)

The production diff, frozen upstream reference, regression coverage and raw benchmark statistics were independently rechecked. No actionable issue was found in the reviewed scope; the production change was kept minimal without additional refactoring. The published tree af19f1ba1c4c899eeaede44a7bf0ded7b95dcace matches the locally tested tree.

The full Linux Release diagnostic build and all 7 CTests passed again. Native Windows/macOS, licensed Cubism rendering and actual GPU/window behavior remain unverified.

CI has not passed: the upstream workflow run for head c1316baf5a02f38d57c83cf14fe5ce41c50c264c is action_required with zero executed jobs and zero check runs. A repository maintainer must approve the workflow before cross-platform CI can execute. Ready for review does not mean ready to merge.