perf: add direct runtime dispatch parity harness

[besmpl]

What changed

• add a pinned, paired Ember/Luau steady-state parity harness with OLS slope fitting, retained fingerprinted attempts, quiet-runner gating, and fail-closed lifecycle self-tests
• split the VM onto a compact production dispatch path with a checked fallback for unsupported or instrumented prototypes
• reduce packed instructions to 12 bytes and deepen numeric-for lowering while preserving coercion, control-flow, and debug behavior
• add mandatory dispatch complexity, opcode, Proto-layout, and numeric parity tests
• record the detailed four-phase implementation plan and run history

Why

The previous comparison included Luau CLI startup and compilation, selected the wrong Phase 1 workload, and could overwrite noisy retries. The VM also paid generic tracing and state-management costs in normal execution. This slice establishes trustworthy measurement and removes the first structural dispatch costs without workload-specific runtime behavior.

Impact

The last valid quiet-run Phase 1 evidence measured arithmetic_for at 1.6865x median and 1.7236x p90 versus pinned Luau 0.728, with the production symbol reduced to 8,272 bytes and stack reservation to 1,680 bytes. Phases 2-4 and final 0.95x/1.00x Scenario parity remain unfinished.

The v5 harness refused live sampling after the machine stayed above its load/CPU limits for 600 seconds; it created no benchmark attempt. This avoids presenting contaminated measurements as current evidence.

Verification

• scripts/check-runtime-parity --self-test
• focused parity-harness tests
• all four required dispatch tests
• concat and open-return allocation regression tests
• go test -vet=off -count=1 ./...
• git diff --check

Local .simplepower journals and ignored raw benchmark artifacts are intentionally not committed.