chore: DWARF / witness mapping discovery for fused output (sub-#130)

[avrabe]

Summary

• Discovery + oracle test for the witness MC/DC integration epic tracked in DWARF / source-map remapping for fused output (witness MC/DC integration) #130.
• meld currently forwards DWARF custom sections byte-for-byte through fusion. Addresses encode offsets into the input code section; the fused code section has different layout, so every preserved address is wrong. With multi-core-module inputs the fused output also carries duplicate .debug_* sections.
• This PR ships only the discovery: a 5-test fixture pinning today's lossy behavior so the planned remap work has a green-to-red signal to flip. No production code changes; defaults unchanged.

Concrete numbers from the new test on tests/wit_bindgen/fixtures/lists.wasm:

DWARF sections at top level of fused module: {
    \".debug_abbrev\": 2,  \".debug_info\": 2,  \".debug_line\": 2,
    \".debug_loc\": 2,     \".debug_ranges\": 2, \".debug_str\": 2,
}
input  code-section length (sum across embedded modules): 231531 bytes
fused  code-section length:                                213242 bytes

Audit findings (committed in the test docstring + filed in #130)

• FuserConfig::default().custom_sections == Merge (meld-core/src/lib.rs:110). DWARF is preserved, not dropped.
• merger.rs:2010-2012 naively concatenates per-module custom_sections with no dedup or rewriting.
• lib.rs::encode_output (line 1345-1356) emits all of them unless Drop is configured.
• No code anywhere in meld-core/src/ parses or rewrites .debug_* content.
• Component-level custom sections are dropped at parse time (parser.rs:1082-1084); only core-module custom sections flow through.

Cross-repo dependency: witness

pulseengine/witness v0.11.x is the consumer that breaks. It uses gimli (crates/witness-core/src/decisions.rs) to build a (code-section byte offset) -> (file, line) map for MC/DC attribution. The DWARF passed through by meld today gives gimli wrong + duplicated input.

Witness is intentionally NOT a meld-core dependency. End-to-end "run witness on a fused module" verification belongs cross-repo (likely in wasm-component-examples release evidence). The test in this PR covers only meld-side invariants.

What this PR contains

• meld-core/tests/dwarf_passthrough.rs — 5 tests, all green:
  • current_default_is_merge_not_drop — pins the default.
  • fixture_carries_dwarf_in_some_embedded_module — sanity-check the fixture still has DWARF.
  • fused_output_inherits_dwarf_byte_for_byte_today — pins that DWARF is forwarded with multiplicity (2 of each section in the fused output).
  • drop_policy_strips_dwarf_completely — pins the user's only escape hatch today (CustomSectionHandling::Drop).
  • dwarf_addresses_in_fused_output_are_known_to_be_wrong — pins the structural "code section length differs" invariant that proves passed-through addresses are invalid.
• TODO comments call out where each assertion flips at Phase 1.5 / Phase 2 / Phase 3.

Phased plan (tracked in #130)

• Phase 1 (this PR): discovery + oracle test.
• Phase 1.5: .debug_*-aware policy distinct from generic custom-section handling. Recommend default = drop for .debug_* until Phase 2 ships, so users at least don't get wrong source attribution.
• Phase 2: real DWARF remap — rewrite .debug_line programs and .debug_info DW_AT_low_pc / DW_AT_high_pc / DW_AT_ranges to the merged code section using the function-body relocation map the merger already builds. Dedup .debug_str. Likely needs a gimli dep on the meld-core side.
• Phase 3: synthesize DIEs for adapter / lift-lower function bodies (or accept the gap, since witness has a strict-per-br_if fallback). Variable-level debug info explicitly out of scope.

Test plan

• cargo test --package meld-core --test dwarf_passthrough — 5 tests green.
• cargo test --package meld-core --test wit_bindgen_runtime — 73 tests green.
• cargo test --package meld-core --lib — 198 unit tests green.
• Pre-commit hooks (cargo-fmt, cargo-clippy, cargo-test) all green.
• Cross-repo verification deferred to Phase 2 (witness end-to-end run on a fused module is the eventual integration oracle).

Refs #130

🤖 Generated with Claude Code