chore(benchmarks): modernize for v1.8 — feature benches, stage output, gate coverage

[DemchaAV]

Summary

The benchmark suite measured only text/table primitives — none of the v1.8
vector features (SVG import, native charts, vector paths, gradients) had a
bench, and the current-speed report blended compose/layout/render into one
number. This branch adds feature-object benches and deterministic probes over
the new surface, splits the current-speed report into per-stage timings with a
readable summary.md, closes a silent gap in the smoke perf gate, and removes
three redundant benches. Everything lives in the separate benchmarks/ module
plus its docs — no library src/main is touched.

Commits (reviewed sequentially)

1. chore(benchmarks): remove three redundant benchmark mains — drop
   FullCvBenchmark (superseded by the JMH TemplateCvJmhBenchmark),
   GraphComposeBenchmark (duplicated CurrentSpeedBenchmark's engine-simple
   scenario), and ScalabilityBenchmark (thread sweep folded into the
   full-profile throughput run, now 1,2,4,8,16); prune the matching
   run-benchmarks.ps1 steps.
2. perf(benchmarks): persist compose/layout/render stages + a run summary —
   CurrentSpeedBenchmark writes a per-scenario stage split (stages[],
   median ms) to the JSON and a stages CSV, plus a human-readable summary.md.
3. perf(benchmarks): diff consumes stages[] and reports added/removed scenarios — BenchmarkDiffTool prints a per-stage delta table and the
   scenario set-change (added/removed) between two runs.
4. perf(benchmarks): add SVG-import feature benches — SvgJmhBenchmark
   (path parse / whole-file icon read / icon→node) + SvgParseAllocProbe.
5. perf(benchmarks): add chart feature benches — ChartJmhBenchmark
   (bar + line + pie render) + ChartAllocProbe (layout-compile allocation).
6. perf(benchmarks): add vector-paint render-operator probe —
   VectorRenderOperatorProbe draws the same paths flat vs. gradient vs.
   translucent and counts the PDF content-stream operators each paint mode emits.
7. bench(jmh): add icon-ramp and mixed v1.8 showcase benches —
   IconRampJmhBenchmark (icon-placement scaling, @Param 8/32/128) and
   MixedShowcaseJmhBenchmark (one document mixing prose, inline sparklines,
   bar + pie charts, SVG icons and a gradient path) as the integration canary.
8. bench(gate): gate the long-token scenario and guard threshold coverage —
   long-token had no SMOKE threshold and silently escaped the gate; add one
   (10.0 ms / 256.0 MB, ~3× its observed ~3.2 ms / ~94 MB) and add
   CurrentSpeedScenarioGateTest, which fails the build if any scenario lacks a
   threshold. Scenario list hoisted to a static SCENARIO_DEFS so the names are
   testable; the six scenarios, order, descriptions and renderers are unchanged.
9. docs(benchmarks): finish the removed-bench cleanup and fix two stale Javadocs — sweep references the removed mains left behind (ab-bench.ps1
   log parsing, performance.md, a merged row in the module README) and correct
   two docs that overstated the code (the SVG-parse fixture has no arc command;
   stages[] is not carried into the median aggregate).
10. docs(changelog): note the v1.8 feature-object benches, stage output, and gate coverage.

Testing

• ./mvnw -B -ntp verify -pl . → BUILD SUCCESS, 1380 tests (canonical suite +
  japicmp + javadoc; the CHANGELOG/docs guards VersionConsistencyGuardTest +
  CanonicalSurfaceGuardTest stay green after the doc edits).
• ./mvnw -B -ntp -f benchmarks/pom.xml verify → BUILD SUCCESS, 30 tests —
  incl. new CurrentSpeedScenarioGateTest and a new BenchmarkDiffToolTest case
  covering added/removed scenarios + stage deltas.
• New JMH benches run end-to-end: IconRamp shows a clean near-linear
  placement ramp (8 → 0.86, 32 → 2.74, 128 → 8.79 ms/op); MixedShowcase
  ~8.7 ms/op. A smoke run with the gate enabled passes with long-token now
  gated. The alloc/operator probes report deterministic counts (chart
  compile-pass ~446.8 KB; gradient paint adds sh+W per shape, alpha adds gs).

Notes for review

• Benchmark-only branch — no library src/main change. The new fixtures
  (SvgBenchmarkFixtures, ChartBenchmarkFixtures) are public only because the
  JMH benches live in the .jmh subpackage; each fixture is reused by both the
  bench and its probe so the two measure identical data.
• stages[] is intentionally not medianed — BenchmarkMedianTool aggregates
  latency/throughput only, so a median-vs-median diff shows no stage deltas; diff a
  single-run pair for stage attribution (noted in the median-tool Javadoc).
• No committed smoke baseline — the verdict gate stays a local same-machine
  A/B tool (a static committed baseline would false-positive on machine variance);
  run-benchmarks.ps1 already skips the verdict step gracefully when none exists.
  The absolute SMOKE gate (now covering all six scenarios) is the CI net.

Lane: test — benchmark tooling + operations docs; no canonical / shared-engine / legacy surface touched.

[DemchaAV]

Closing — opened prematurely. Work continues on the chore/benchmark-modernization branch; the PR to develop is the maintainer's call once the branch is finalized.