perf(runtime): call the guest execute helpers directly instead of compiling snippets

[sd2k]

Summary

The guest's execute path built a _eryx_exec("""<escaped user code>""", False) string and ran it through PyRun_SimpleString, then compiled three more fixed snippets (_eryx_output, _eryx_errors = _eryx_get_output(), _eryx_capture_result(), _eryx_discard_result()) and re-resolved __main__ four times to read the results back. So CPython tokenised an escaped copy of the user's source inside the wrapper and compiled it again inside _eryx_exec, on every request.

This resolves __main__.__dict__ and the four _eryx_* helper function objects once, at the end of initialize_python() (owned references in a OnceLock; they live in linear memory and survive the pre-init snapshot like everything else), and execute_python now:

• calls _eryx_exec(code, trace_enabled) with PyObject_Call, passing the source as a str object — no wrapper string, no escaping, no second parse;
• reads (stdout, stderr) from the tuple _eryx_get_output() returns;
• calls _eryx_capture_result() / _eryx_discard_result() directly and reads _eryx_result / _eryx_result_error with two dict lookups.

Error semantics are unchanged: on an exception PyErr_Print() runs before the streams are restored, so the traceback lands in the captured stderr exactly as PyRun_SimpleString produced it (the contract from #239). The pending/async path is untouched. If the helpers are missing from __main__ for any reason the old PyRun_SimpleString path is used, so a broken infrastructure snippet degrades rather than fails.

Stacked on #416 (perf/guest-callback-setup-cache) because that is where the session path stopped being dominated by callback re-setup; the python.rs changes are disjoint from it. prebuilt/liberyx_runtime.so.zst is rebuilt from this branch, so the usual conflict-on-merge applies to that file.

Benchmarks

Criterion (cargo bench --package eryx --features embedded --bench execution), stock wasmtime 48.0.1, Ryzen 9 7950X, this branch vs its base #416:

Benchmark	#416	this PR	Δ
session_execution/pass	187.3 µs	85.9 µs	−54%
stateless_execution/pass (3 callbacks)	4.90 ms	4.79 ms	−2% (dominated by per-instance callback setup)
session_creation/new	449 µs	441 µs	— (not touched)

Spike harness (profile_stateless, 2000 executions, median of 3; measured on top of the wasmtime pagemap/funcref fixes from the cold-start spike so instantiation cost does not hide the guest cost):

Path	before	after	Δ
warm instance, 2 ms gap (request-like), trace collection on	282 µs	145 µs	−48%
warm instance, trace collection off	196 µs	82 µs	−58%
cold instance	497 µs	400 µs	−19%
cold render workload (json + string.Template), trace off	903 µs	815 µs	−10%

In the warm-path profile the parser/compiler symbols (_PyPegen_*, _PyCfg_*, _PyAssemble_*, tok_get_*) drop from 6% to 1.8%; what remains is compile() of the user's own source.

Testing

• cargo nextest run --workspace --features embedded --cargo-profile release: 629 passed, 0 failed.
• cargo clippy --workspace --all-targets --all-features -- -D warnings: clean.

Not included

• Skipping the list-callbacks host call on the execute path (still made every execute; cheap next to the above).

🤖 Generated with Claude Code

[sd2k]

Rebased onto main (dropping the now-merged #410/#416 commits) and rebuilt prebuilt/liberyx_{runtime,bindings}.so.zst against the wasm-tools 0.259 toolchain from #414, so the committed guest artifact matches what the current tree builds. Verified the rebuilt .so actually carries the new code (the fallback path means a stale artifact would pass tests silently) and that the fast path is live: session_execution/pass 48.7 µs with ERYX_PROFILE_TRACE=0. 629 tests, clippy -D warnings and fmt all clean locally.

[github-actions]

🌐 Demo preview: https://perf-guest-direct-execute.eryx-bvy.pages.dev
📌 This deploy: https://96ec3914.eryx-bvy.pages.dev