fix(wasm): module-level member writes persist on the web target (#5016)

[proggeramlug]

Summary

Fixes #5016. On the web/wasm target, a write to a module-level array element or object property (A[0] = ..., o.prop = ...) was silently dropped — subsequent reads returned the initializer. Native always worked. This broke the documented module-scope mutable-state pattern (const S = [...]; mutate S[i]) that Perry's own pitfalls guidance recommends, so games/UIs built for web never updated state (discovered shipping Bloom Jump to web). Regression of #1993.

Root cause

The HIR lowers both member-write forms — obj.prop = v and obj[i] = v — to a single Expr::PutValueSet node. The WASM backend (perry-codegen-wasm) never had a match arm for it, so every such assignment fell through emit_expr's _ => TAG_UNDEFINED catch-all and the store was never emitted. The still-present IndexSet/PropertySet arms had become dead code once lowering switched to PutValueSet — which is why the earlier #1993 fix appeared to regress.

Fix

Handle PutValueSet in the WASM backend, mirroring the pre-PutValueSet lowering split so behavior matches the existing arms:

• string-literal key (obj.prop = v) → class_set_field (so class getters/setters still fire)
• any other key (obj[i] = v) → object_set_dynamic
• both return the assigned RHS value (assignment-expression semantics)

Companion expression walkers updated in tandem so the node is fully supported: string_collection.rs (intern key string + value literals), closures.rs (collect closures used as value/key/target), js_fallback.rs (emit equivalent JS assignment).

Verification

• Issue repro and variants (in-function, top-level, through-parameter, object property, class setter) now match native on web.
• WASM target suite: 17 → 21 passing, 0 regressions. The four newly-green tests (05_objects_arrays, 07_classes, 08_getters_setters, 21_class_method_i64) were all broken by this same dropped-write bug. The 6 remaining failures are pre-existing and unrelated (higher-order array methods, map/set, regex/date, splice, ffi imports).
• Added tests/wasm/26_module_array_element_write.ts regression guard.

🤖 Generated with Claude Code

[proggeramlug]

CI status

This branch is rebased on latest main (v0.5.1161). Beyond the #5016 wasm fix, it also clears two pre-existing cargo-test blockers on main that were surfacing one-per-run:

1. fix(stream) — readable.read() with no size now follows Node's howMuchToRead(NaN) (head-chunk only while flowing; full buffer while paused) instead of always returning the head chunk, fixing the node_stream unshift unit tests regressed by feat(runtime): node:stream + node:stream/web (Readable/Writable/Transform + WHATWG) (#1545) #5017.
2. fix(zlib) — the API manifest declares deflateRawSync(buf, options?) to match the dispatch arity fix(stdlib): #4917 — real semantics (or loud warn) for silently-ignored adapter options and no-ops #4996 introduced, fixing manifest_param_counts_match_dispatch_table.

Green: lint, api-docs-drift, security-audit, harmonyos-smoke.

Still red — pre-existing main breakage, NOT introduced by this PR (diff is wasm-codegen + one stream fn + one manifest line; touches no GC or native codegen):

• cargo-test → gc_write_barrier_stress (missing old→young remembered-set edges → segfault) — tracked in CI: gc_write_barrier_stress red on main (missing old→young remembered-set edges → segfault) #5029.
• compiler-output-regression → h1_buffer_alias_negative native workload — tracked in CI: compiler-output-regression red on main (h1_buffer_alias_negative native-region-proof workload) #5030.

Both reproduce on a clean main and need dedicated GC / native-codegen context; fixing them here would be out-of-scope, risky surgery.