Add reproducible local benchmark and Unity compatibility matrix harness

[wallstop]

Context

The modernization plan now has focused correctness coverage, but the release gate still lacks repeatable performance and cross-version evidence. Current validation is local Unity 6000.4.6f1 on the confirmed host; Unity 2021.3/2022.3, other local platforms, and measured Play-entry budgets remain unverified.

Scope

Add a disposable local driver under scripts/ that accepts explicit host-project, Unity-version, fixture-size, suite, and output-directory arguments. It must support the existing Unity MCP bridge and direct host Unity execution with explicit path translation.

The fixture and measurements should cover:

• deterministic 100, 1,000, 10,000, and 50,000 main ScriptableObject datasets
• BaseDataObject and ordinary ScriptableObject controls, labels, descriptions, nested fields, shared references, duplicate short type names, custom ordering, and saved selections
• shell construction, first visible rows, selection readiness, complete indexing, search, selection, filtering, scrolling, type switching, refresh, allocations, retained references, and realized rows
• cold import, warm restart, warm in-session, incremental edits, and Play entry with the package absent, closed, open/idle, and open/indexing
• machine-readable raw samples and readable comparisons recording revision, fixture, Unity, OS, CPU, memory, storage, graphics, DPI, theme, window size, and reload settings

Acceptance criteria

• Generated fixture projects, datasets, caches, and results remain outside the distributed package.
• Timing excludes process launch, MCP transport, and command dispatch; Unity measurements use monotonic timers and focused profiler markers.
• Warm UI and Play-entry scenarios use five warm-ups and at least 30 alternating matched repetitions; expensive restart/compile scenarios use at least five samples and report all samples.
• The driver verifies fixture counts/content before timing and restores any changed project settings, including after failures.
• A compatibility report covers Unity 2021.3, 2022.3, and the current host editor, with player compile/build smoke results for ordinary and BaseDataObject assets.
• The report identifies whether Play-entry p95 overhead meets the 20 ms target and separates window-open from window-closed cases.

Related: #19

[wallstop]

Implemented in PR #19 commit 6fe4cfc:

• Added scripts/benchmark/benchmark_driver.py with direct Unity and official MCP modes, explicit host/container path mapping, disposable deterministic 100/1,000/10,000/50,000 fixtures, JSON output, and Markdown comparison reports.
• Unity-side timing uses Stopwatch plus named ProfilerMarkers; Play-entry uses 5 warm-ups and 30 alternating samples for open and closed cases.
• Real Unity 6000.4.6f1 MCP validation verified the 100-asset fixture and cleanup.
• The real 60-sample Play-entry run measured open p95 281.9102 ms and closed p95 242.2834 ms against the agreed <=20 ms target; the report records playEntryTargetMet=false.

Remaining issue #21 work is intentionally explicit: larger fixture runs, package-absent controls, cold/warm/index/search/render/memory metrics, compatibility/player matrix, and broader comparison baselines.