Summary

The final piece of DWARF Phase 2 (#143): DwarfHandling::Remap turns the AddressRemap engine (v0.18.0) into a working, debug-info-preserving fusion path. It reads an input core module's .debug_* sections, translates every code address to the fused code section, and re-serializes a single remapped DWARF set via gimli::write::Dwarf::from. Exposed as meld fuse --dwarf remap.

This completes the increment arc:

Design — three decisions that de-risk it

1. Post-hoc parallel operator walk. The instruction offset map is recovered by walking the input and final-output operator streams in lockstep, not captured during the merge. The merge re-rewrites bodies after adapter wiring (lib.rs:694), so a captured map would go stale; the post-hoc walk reflects whatever rewriting actually happened and threads no state through the hot path. A per-function operator-count or locals-prefix mismatch (e.g. memory-rebasing inserted scratch ops) aborts the remap.

2. Correct-or-strip. gimli::write::Dwarf::from is all-or-nothing on addresses — a None from convert_address aborts the whole conversion. That's used as the safety gate: only the structurally-invariant code-section base (address 0) is special-cased; any other unmapped address fails the conversion and falls back to stripping. Never emit a wrong address.

3. Single-DWARF-source scope. DWARF is per-input-core-module but the output has one code section. Merging N independent DWARF unit sets is a separate fidelity problem — deferred. Exactly one DWARF-bearing source → full remap; more than one → strip + warn; zero → no-op. The common single-component case (main module carries DWARF, hand-written adapter modules don't) is fully served.

A three-pass encode keeps the remapped .debug_* inside the attestation/provenance-hashed bytes (trailing custom sections don't shift code offsets, so the remap built from pass A is valid for the final output).

Verification

• LS-D-1 (new, approved): wrong remapped DWARF address → de-grounded downstream coverage/breakpoints. Gated by dwarf::tests::ls_d_1_remap_translates_low_pc — a full gimli read→convert→write→read oracle that builds real input DWARF, remaps a subprogram's low_pc, re-parses the output DWARF, and asserts the address was actually translated.
• Parallel-walk unit tests: identity round-trip from real wasm bytes + operator-count-mismatch abort.
• Multi-source strip-fallback integration test on lists.wasm.
• The six translate_* unit tests (3a) pin the remap math.
• tools/run_ls_verification.py: [ OK ] LS-D-1 (1 pass).

Residual (documented in LS-D-1): DW_AT_high_pc encoded as a length (DW_FORM_data*, the common Rust/LLVM encoding) is copied verbatim — gimli treats it as a constant, not an address — so a function's reported byte length may be off by intra-function LEB drift. low_pc and the line-number program (what debuggers and pulseengine/witness use) are correct.

Notes

• Adds the gimli dependency.
• dwarf.rs Tier-5 registration is a separate follow-up PR (the claude-code-action byte-identical-workflow constraint forbids bundling a mythos-auto.yml edit with code changes).

🤖 Generated with Claude Code