Preserve instruction origins in rewritten DWARF line tables

[cpunion]

Depends on #9196 (which depends on #9192). The standalone diff contains only instruction-origin handling. Start with the implementation and tests; five mechanically regenerated snapshots are in a separate commit. GitHub's main-based comparison still includes the unmerged prerequisites.

Fix

Asyncify's inserted rewind/unwind code must not inherit a neighboring user's source line. Surviving instructions must also retain their location when the instruction starting the original line-table row is removed.

Record origins while writing actual instructions, including instructions introduced by the binary/Stack IR writer. Treat input line rows as address intervals: recover original instructions' file/line/column/discriminator, give inserted code line = 0, column = 0, is_stmt = false, and restore source attribution when original code resumes.

Emit rows directly in writer order, coalescing unchanged attribution once. Only function-header/end-sequence records are retained separately; there is no second IR walk, full output-row map, sequence-normalization pass, or second coalescing pass. Functions without input line metadata, including Asyncify helpers, are not assigned fabricated source DWARF.

Generic line-program compression and discriminator serialization are now isolated in #9196. This PR does not handle local re-indexing, generated helper DIEs, or TinyGo's missing metadata, and does not fully resolve tinygo-org/tinygo#5736.

Validation and cost

• Imported-call tests cover deleted row starts, Asyncify with/without -O1 and Stack IR, original/generated instruction boundaries, discriminator and row flags, unrelated functions without line metadata, and a second roundtrip. A separate loop fixture distinguishes original traps from writer-inserted unreachable instructions. Both regression tests fail on Correct and compact DWARF line-program encoding #9196 and pass here.
• All 12 DWARF unit tests, 95 Python unit tests, and the full wasm-opt --no-torture suite pass locally. C++: 437 passed, one existing macOS size-test skip. Ruff and changed-line clang-format checks pass. V8-specific execution checks are left to CI.
• A representative TinyGo goroutine/JSON input passes LLVM 22 DWARF verification for roundtrip, -O1, and -O1 --asyncify; the Asyncify output runs under Wasmtime and prints {"name":"dwarf","value":42}. Stripped outputs remain byte-identical to both prerequisites. The full Netbird application has not been tested.

Mode	#9196 module bytes	This PR module bytes	#9196 time	This PR time
Roundtrip	1,893,879	1,895,081	167 ms	171 ms
-O1	1,872,431	1,937,517	212 ms	201 ms
-O1 --asyncify	2,146,432	2,460,857	423 ms	458 ms

Times are nine-run interleaved local medians after warm-up; background system activity makes small differences inconclusive. These are debug-processing costs, not runtime performance. Asyncify executable code stays exactly 671,942 bytes; the extra debug rows increase file size and peak RSS (approximately 150 → 185 MiB, three-run medians). This simplification reduces implementation complexity but does not establish a peak-memory improvement.

This is an original contribution, developed with AI assistance.

[tlively]

@kripken, do you have more context for this?

[kripken]

No, I have no more context here.

@cpunion it would be good to clarify what is fixed here with a small example.

Also the overview could be more specific: "Preserve instruction origins in rewritten DWARF line tables" is something we were already doing, or at least trying to.

[tlively]

Asyncify's inserted rewind/unwind code must not inherit a neighboring user's source line.

Why is this?