feat(option-A): per-component handle tables + bridging (3/3 trio passes)

[avrabe]

Summary

Implements Option A from issue #107 — per-component handle tables with cross-component bridging trampolines. Builds on PR #108's path B foundation. All three trio fixtures now pass at runtime, closing issue #75.

Phase 1 (commit 0632ec2)

Broaden reexporter_resources to also include the resource's DEFINER component. With resource_with_lists, this allocates 2 ht entries (was 1): one for the leaf definer, one for the intermediate re-exporter.

Phase 3 (commit 8437e3f)

Insert bridging trampolines in fact.rs adapter:

• own result transfer when both caller and callee have ht: callee.ht_rep + caller.ht_new translates returned handles into caller's namespace.
• borrow param transfer when caller has its OWN ht: caller.ht_rep + callee.ht_new translates outgoing handles into callee's namespace.

Phase 4 — trio close-out (commit 9149573)

Three layered fixes for the remaining 2 of 3 trio runtime failures:

1. fact.rs — borrow-rep discriminator (both 2-comp and 3-comp branches). Method-like exports ([method]/[static]/[constructor]) on a locally-defined resource expect the REP as arg0 (wit-bindgen's _export_*_cabi calls ThingBorrow::lift(arg0) which derefs as *mut _ThingRep<T>). Top-level functions taking borrow<T> on a used resource expect a HANDLE (cabi calls Float::from_handle(arg0 as u32)). Suppress callee.ht_new for method-like calls; emit only caller.ht_rep so the rep flows through unchanged. Without this, the freshly-minted slot address gets passed as the rep, the deref reads 4 bytes at the slot (the just-stored rep), and Option's discriminant byte is the low byte of that rep — 0 for typical aligned box pointers → Option::unwrap on None.

2. merger.rs — dtor suppression for re-exporter ht_drop. Phase 1's per-component HTs for definers store foreign reps placed there by own bridges (intermediate's HT contains leaf box pointers). The dtor (<iface>#[dtor]<rn>) blindly casts every stored value as *mut _ThingRep<LocalT> and Box::from_raw drops it; for foreign reps this misinterprets memory and triggers re-entrant drops via the wit-bindgen Resource::drop impl, producing unbounded recursion. Skip the dtor when the component is a re-exporter — the standard Box-pattern wit-bindgen design assumes the rep is owned by the component whose ht stores it, but Phase 1 broke that invariant.

3. resolver.rs — sort reexporter_components and reexporter_resources before storing on the graph. They were collected from HashSet, whose iteration order is non-deterministic. Downstream HT allocation and the wrapper alias-fallback both make first-match decisions on this order, so the same fixture would sometimes wire [resource-drop] for the runner to leaf's ht_drop and sometimes intermediate's, producing the "passes manually, fails in cargo test" flakiness.

Status

	Result
73-test wit-bindgen suite	✅ 73/73 (0 regressions)
resource_floats runtime	✅ PASSES
resource_with_lists runtime	✅ PASSES
resource-import-and-export runtime	✅ PASSES

All three trio fixtures promoted from fuse_only_test! to runtime_test!. Verified stable across 5/5 consecutive cargo test runs after the determinism fix.

Related

• Builds on feat: per-resource handle tables + alias-fallback (1/3 trio runtime passes) #108 (path B + alias-fallback)
• Closes the trio runtime failures from Path B partial: per-resource handle table routing complete; outer-boundary handle leak remains #107
• Closes Promote resource_floats and resource_with_lists to runtime tests #75 (trio runtime fixtures)
• Part of feat: per-component resource handle tables for 3-component chains #69 (Epic: per-component resource handle tables)
• Part of 3-component re-exporter resource chains: implement shim module approach #92, Epic: static fusion of re-exporter resource chains (ADR-0) #106 (parent epics)

Test plan

• cargo test --release --test wit_bindgen_runtime — 73/73
• resource_floats runtime passes
• resource_with_lists runtime passes
• resource-import-and-export runtime passes
• Stable across 5 consecutive runs

Branch

feat/per-component-ht-bridging head 9149573 — 3 commits on top of feat/conditional-resource-typing.

[temper-pulseengine]

AI Code Review for PR #109

pulseengine/meld:feat/per-component-ht-bridging → pulseengine/meld:main

This pull request introduces a new feature in the Meld tool, which is a static component fusion engine for WebAssembly. The main changes are:

1. Opaque Resources: A new configuration option --opaque-rep has been added to mark resources as opaque-rep. This means that these resources have their representation treated as a u32 rather than a boxed pointer, and they skip the per-resource handle table layer in Meld.

2. Resource Type Management: Opaque-rep resources are routed differently from standard Box-pattern resources:

   • The merger.rs::allocate_handle_tables function skips them since their reps are already valid integer handles.
   • The component_wrap.rs::local_resource_types function keys them by (component_idx, resource_name) rather than resource_name alone. This ensures that each component owns its own wasmtime resource type, matching the un-fused composition's semantics.

3. Command Line Interface: The Meld CLI now supports parsing and handling the new --opaque-rep option. It also prints a list of opaque-rep resources if specified.

4. Documentation Update: The documentation has been updated to reflect these changes, including examples and explanations for the new configuration option.

Overall, this feature enhances Meld's ability to handle complex WebAssembly component compositions while maintaining performance and security.

---

This review was generated by a local AI model. It is advisory only and may contain inaccuracies.

Reviewed at 8437e3f