perf: reduce runtime overhead and memory usage

[owenthcarey]

This PR reduces runtime allocation, copying, and dispatch overhead across JSON, strings, collections, datetime, serialization, and native/JIT calls. It also adds regression coverage and keeps generated research artifacts out of the source tree.

Changes

• Decode and encode JSON with fewer intermediate buffers; avoid whole-input string scans for bounded prefix/suffix checks and splitting.
• Add native fast paths for deque operations, datetime arithmetic and formatting, numeric reductions, range collection, and supported pickle values, with guarded fallback paths.
• Reduce Object and DictKey from 24 to 16 bytes on ARM64 through shared value storage. Use lazy instance metadata, compact slots, shared slot names, and reusable collector scratch space.
• Improve JIT constants, scalar calls/results, list iteration/construction, native entry, and exception exits. Bound retained native-call scratch capacity while preserving callback, deoptimization, weak-reference, and C API lifetimes.
• Add differential probes and regression tests, remove generated logs/results/source snapshots from the final tree, and enforce repository artifact rules in CI.
• Fix Windows compilation after the storage refactor by sharing lazy DLL-directory instance dictionaries and gating a Unix-only import.

Measured improvements

These are recorded macOS ARM64 measurements against each change's own predecessor, not a fresh comparison of the final PR against main. Results from different stages must not be multiplied into a cumulative speedup.

Workload or change	Recorded improvement
JSON roundtrip	73.0% less JIT workload time; 72.8% less interpreter workload time
JSON record encoding / decoding	4.07x / 3.53x faster
JSON numeric-array decoding	11.85x faster
Large ASCII / Unicode JSON decoding	88.4% / 61.1% lower peak RSS
Deque fixture	35.8% less workload time, 34.8% less process CPU time, 5.2% lower peak RSS
Datetime fixture	47.6% less workload time, 47.1% less process CPU time, 19.1% lower peak RSS
Shared value storage	2.34% lower geometric-mean peak RSS across 24 processes; every process improved
Public time.isoformat	About 37% to 39% less execution time

The initial JSON pass reduced the 23-workload suite's geometric-mean workload time by 6%. The subsequent deque/datetime pass reduced JIT workload time by another 5.3% against its immediate predecessor. The JSON report, deque/datetime report, and runtime report document methods, tradeoffs, and reproduction commands.

Limits: WeavePy does not yet outperform CPython across all workloads or metrics. Some short allocations, calls, fallback paths, and startup measurements regressed. The final scratch-capacity limit has no qualified timing samples. Later archived JIT/datetime experiments are not part of the active implementation and their gains are not claimed here.

Validation

• The recorded runtime checkpoint passed 353 VM tests, 156 C API tests, 99 targeted checks, 44 independent CPython result checks, and 275 compatibility checks, plus formatting, strict lint, and no-JIT checks.
• The Windows CI fix passes local formatting, the repository artifact check, all 11 repository-tool tests, whitespace validation, and a Windows GNU cross-check of the VM and its tests with all features and Rust warnings denied. The Windows CI log confirms that the DLL-directory creation, closing, and context-manager regression passes.
• CI for the updated sources passed all 28 jobs for 9a10bae5, including platform tests, bundled and free-threaded regressions, distribution checks, benchmark gates, and all nine ecosystem shards.

Generated benchmark evidence remains available locally and in checkpoint history. Required Unicode/codec data, vendored build inputs, CI baselines, reusable tools, and regression fixtures remain tracked.