perf(aot): inline/cold-outline three per-step hot leaves (~6% less CPU work)

[pyrex41]

What

A fresh --kernel-tests leaf profile (post split-TLS / intern-cache round) flagged three small functions sitting as un-inlined call frames in the per-step hot loop. Each had a tiny hot body but a cold error path (format! / ShenError::cancelled string-building) that bloated it and blocked LLVM from folding it into its AOT call sites. Split each into a tiny #[inline] hot path + a #[cold] #[inline(never)] error constructor:

leaf	role	self-samples (before → after)
is_truthy	the AOT if predicate	55 → 7
charge_step	per-step budget/deadline check	45 → 0 (inlined into eval_in)
make_aot_closure / global_value / fn_value	AOT lambda/value/fn	re-probed intern HashMap → now intern_static (pointer cache)

The last three took &str and re-probed the intern HashMap on every AOT lambda/value/fn evaluation; the AOT call-target path (apply_named/apply_direct) already used the pointer-cached intern_static, so this just extends the same fast path.

Measurement

~6.2% less CPU work — paired user-CPU min-of-13, B < A in ~9/11 runs. The machine was loaded (a video call); wall-clock minima sat at ~2× the clean floor and were unusable, so user-CPU time (which doesn't inflate when other processes steal the core) was the contention-robust proxy. Re-run scripts/cross-port-bench.sh quiet for a clean wall-clock confirmation.

The work doesn't vanish — it CSE's into eval_in and the call sites — but the per-step call/return overhead and the cold-blob bloat are gone.

Correctness

• 134/0 across tree-walk, SHEN_RUST_VM=1, SHEN_RUST_GC=1, and --served.
• clippy + fmt clean.
• charge_step behaviour identical — the sticky step-budget-exhaustion semantics live unchanged in the outlined charge_step_limited; the fast path early-returns only when both budget and deadline are unset (the default).

Also

• scripts/cross-port-bench-4way.sh — extends the headline harness to LuaJIT and PUC Lua (rust ~2× faster than LuaJIT, ~4.5× faster than PUC Lua on this suite).
• BENCHMARKS.md records the 4-way field and the LuaJIT FNEW/UCLO trace-abort finding (filed upstream as shen-lua#27).
• PERFORMANCE.md adds this as round 5 of the gap-closing log.

🤖 Generated with Claude Code

[pyrex41]

Review pass (cursor-agent, GPT-5.5-high) — addressed in 2383282

Ran an independent read-only review with gpt-5.5-high. Verdict was REQUEST-CHANGES on one real bug; it confirmed the runtime change itself is correct (it even ran cargo test -p shen-rust --test budget_cancel → 5/5, validating the sticky step-budget path).

Fixed:

• HIGH — cross-port-bench-4way.sh captured empty timings. /usr/bin/time's stderr was redirected to /dev/null inside the timed group before awk could read it. Restructured to silence the program's own output inside an inner sh -c; verified it now emits real numbers. Dropped the dead bench()/run_* helpers it flagged.
• Correctness fallout from that fix: the working script contradicted a claim I'd written in BENCHMARKS.md ("rust ~2× faster than LuaJIT") — that number compared rust's internal eval timer to a contended LuaJIT wall-time. Under the consistent harness rust and LuaJIT are ~tied (~2.5 s). Corrected the section (firm: cl fastest, PUC Lua slowest; rust-vs-LuaJIT unresolved → re-run quiet). The load-independent FNEW/UCLO trace-abort finding stays, reframed accurately.
• LOW — intern_static keyed on address only. A future &'static str caller passing a prefix slice of another literal would collide. Now keyed on (addr, len).

Confirmed correct, no change needed: charge_step fast-path preserves sticky exhaustion + deadline semantics; the &str → &'static str signature changes are satisfied by all AOT-emitted literals (a non-static caller would fail to compile, not silently misbehave); codegen-units = 1 is not required for cache correctness.

Re-verified: 134/0 (tree-walk + VM), fmt + clippy clean, intern unit tests pass.

Note on the headline ~6% number: that was measured with a different, correct harness (program stdout-only redirect + user-CPU time as the contention-robust proxy), not the buggy script — so it stands. A clean-machine wall-clock re-run is still the right confirmation.