Perf baseline vs the jco legs (and consumer case-budget fit)

[lann]

Promised by the C3 milestone row and never measured: a coarse comparison of the deltic legs against the jco legs before consumer migration — websocket suite wall-time, webcrypto ops/sec, iroh handshake latency (where the ~5x bounded-polling removal is the expected win; the v1 interpreter host boundary is the expected risk). The consumer suites carry per-case budgets/timeouts; we should know we fit them before their CI discovers it. Feeds #14; the generated-JS executor (#8) is the escape hatch if a gap shows.

[lann]

Operator ordering rule (2026-08-10): this comparison runs before any repo's jco legs are deleted. tls's removal was reverted accordingly (polymorph-tls#41) — so every venue is live again: websocket, tls, webrtc-datachannels, iroh all have jco and deltic legs side by side on main (iroh's removal is parked as draft polymorph-iroh#39), and webcrypto retains jco indefinitely (#14 ruling). Measurement notes per repo: the driver-ct suites share per-case wall-clock in the L4 JSONL (duration-ms), so a first-order comparison may need no new harness — aggregate the standing results per target; iroh additionally has scripts/bench.sh (handshake/roundtrip/bulk medians, wasmtime-only today) as the shape to extend if a deeper leg is wanted.

[lann]

Both halves answered; the instrument is landed. Closing.

Consumer case-budget fit: every deltic leg across the family passes the same per-case budgets and wall bounds as the jco legs it sits beside (all matrices green; the one accommodation anywhere is webcrypto's declared instance-reuse for its 19k-case corpus, a renderer memory ceiling rather than a time budget).

Lane-level baseline (browser legs, same box/Chromium — recorded on #14): tls ~parity (7-case suite, fetch-dominated); websocket 32.8→16.5 s (~2×); webcrypto 155→51 s (~3×); webrtc 38.2→27.2 s (~1.4×, ICE/DTLS-bound). Directionally uniform in deltic's favor; confounded by lifecycle/policy differences, which is why the boundary microbench exists.

Boundary-level baseline (bench/boundary, #56 — same Node binary for every lane, harness-free): deltic's callback ABI sustains 0.4–1.2 M async import crossings/s; jco's async import path costs ~3 ms/call flat (its sync path is healthy at ~300–350 k/s — the cost is the async task loop, i.e. the lann/jco#11 machinery, and it quantitatively explains both the lane-time gaps above and polymorph-iroh's 5 ms polling workaround). Sync-lowered calls: deltic ~2–5× faster. Found along the way: #54 (per-element list<u8> lift, ~22 MB/s host→guest vs ~6.4 GB/s guest→host) and a ~2× jspi-vs-callback deferred-settlement delta recorded for #8's ledger.

just bench-boundary re-measures the current tree (the recv@1200 rows are #54's regression sentinel); with-jco keeps the incumbent lane runnable until the jco trees retire.