kilnd: E03ED 'Failed to load module' on witness-instrumented cores (real-app scale) — WASI itself works

[avrabe]

Reproduced the full witness MC/DC path against kilnd (WIP build, 2026-07-08) and isolated the blocker. It is not a WASI or execution bug — WASI runs fine — it's the loader/decoder choking on what witness instrumentation adds at real-app scale.

Repro

wasi component (bin → wasm32-wasip2, 13 wasi imports)
  → meld fuse            → single core (28 globals, 65 KB) — WASI imports inlined
  → witness instrument   → core (1413 branches, 4267 globals, 2 memories, 258 KB)
  → kilnd WITNESS_MODULE=… WITNESS_OUTPUT=…

kilnd fails at load:

✓ WASI enabled (Preview2)  ✓ Component model enabled
is_component=false → execute_traditional_module → Memory OK → Engine OK
→ WASI dispatcher connected → ✗ [Runtime][E03ED] Failed to load module

Isolation (the useful part)

artifact	globals	memories	size	kilnd
un-instrumented fused WASI core	28	1	65 KB	✅ runs — "execution completed successfully"
synthetic: 9000 globals + _start	9000	0	116 KB	✅ runs
synthetic: 2 memory declarations	0	2	tiny	✅ loads (reaches "no entry point")
witness-instrumented fused core	4267	2	258 KB	❌ E03ED at load

So, ruled out: WASI (un-instrumented WASI core runs), witness output (wasm-tools validate accepts it), raw global count (9000 OK), module size (116 KB OK), and multi-memory declaration count (2 OK).

Leading candidate (couldn't fully isolate — wasm-tools' text parser fights the multi-memory-store WAT)

The one structural thing the instrumented core adds that I could not rule out: witness's per-row instrumentation exports a second memory (__witness_trace, memory index 1) and helper functions (__witness_trace_record / _reset / _row_marker) that read/write memory index 1. A 2-memory declaration loads, but using index 1 is the untested delta. Strong suspicion: multi-memory use (non-zero memidx in load/store, or trace-memory instantiation) isn't handled at load/instantiation.

Other un-ruled-out possibilities the decoder owner can check faster than me: passive data/elem from fusion, the added exported functions, or a per-branch construct at count.

Why it matters

This is the single blocker for meld fuse → witness instrument → witness run --harness kilnd on real WASI components (the intended MC/DC path). kiln#343's example "worked" only because it was ~2 globals / 1 memory. Repro module saved at /private/tmp/wasimc-inst.wasm (258 KB); happy to attach or regenerate. Filed from the witness side; this looks like it lives in the kiln-decoder / memory-system area you're already in.