perf(executor): per-ExecutionCtx ArrayKernels snapshot

[lukekim]

Summary

Resolves ArrayKernels once at ExecutionCtx construction into a crate-private KernelSnapshot (a single ArcSwapMap::load_full of the execute-parent kernel map), instead of cloning the session and probing the sharded session-variable DashMap on every array node in the execute_until / single-step execution paths.

Why

• Removes a per-array-node session clone + DashMap shard RwLock probe from the hot execution loop.
• Stops holding the session-variable read guard across kernel invocation (previously a plugin kernel touching the session registry could contend/deadlock on the same shard).

Semantics

The registry is session-scoped and mutable via its public register_* methods. An ExecutionCtx sees a point-in-time snapshot taken at construction; later registrations are picked up by the next context (contexts are created per evaluation). Kernel lookup order is unchanged: registered plugin kernels are tried before static execute_parent kernels, with the same (parent, child) hashing.

Adds a pub(crate) ArcSwapMap::load_full accessor (snapshot-that-outlives-the-call, complementing read). KernelSnapshot and ArrayKernels::snapshot() are pub(crate) — no new public API.

Testing

• cargo nextest run -p vortex-array — 2893 passed (includes struct_cast_execute_parent_uses_session_plugin, covering the register → snapshot → execute path).
• cargo clippy --all-targets -p vortex-array — clean on the spiceai-53 variant of this change; develop port re-verified via the full test build.

[codspeed-hq]

Merging this PR will not alter performance

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚡ 52 improved benchmarks
❌ 62 regressed benchmarks
✅ 1423 untouched benchmarks
⏩ 10 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	chunked_bool_canonical_into[(1000, 10)]	20.7 µs	36.3 µs	-43.08%
❌	Simulation	compare[48]	213 µs	300.4 µs	-29.11%
❌	Simulation	compare[50]	227.7 µs	318.9 µs	-28.58%
❌	Simulation	compare[49]	228.2 µs	317.5 µs	-28.13%
❌	Simulation	compare[46]	218.5 µs	302.2 µs	-27.69%
❌	Simulation	compare[47]	223.5 µs	309.1 µs	-27.68%
❌	Simulation	compare[44]	212.2 µs	292.1 µs	-27.37%
❌	Simulation	compare[45]	218.9 µs	300.7 µs	-27.21%
❌	Simulation	compare[40]	195.6 µs	267.3 µs	-26.82%
❌	Simulation	compare[43]	214.2 µs	292.3 µs	-26.71%
❌	Simulation	compare[42]	209.4 µs	285.6 µs	-26.68%
❌	Simulation	compare[41]	209.3 µs	283.8 µs	-26.22%
❌	Simulation	compare[39]	204.7 µs	274.5 µs	-25.43%
❌	Simulation	compare[38]	200.3 µs	268.3 µs	-25.33%
❌	Simulation	compare[32]	173.4 µs	231 µs	-24.96%
❌	Simulation	compare[36]	194 µs	258.2 µs	-24.86%
❌	Simulation	compare[37]	200.1 µs	266.2 µs	-24.83%
❌	Simulation	compare[35]	195.4 µs	257.7 µs	-24.19%
❌	Simulation	compare[34]	191.1 µs	251.6 µs	-24.03%
❌	Simulation	compare[33]	190.6 µs	250.1 µs	-23.78%
...	...	...	...	...	...

ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing spiceai:lukim/exec-kernel-snapshot-develop (2820b17) with develop (d0013ff)}

Footnotes

1. 10 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩