2026-08-19 - The encoding table stops being a guess

Done. Four things, in the order they unblocked each other.

• Differential harness. orbistoun-gen fixtures compiles shaders from source this project wrote and turns LLVM's disassembly into committed fixtures; tests/differential.rs checks our instruction boundaries against them.
• Registers. orbistoun-gpu::registers extracts register writes from a walked submission and reassembles shader addresses from them - the link that turns two disconnected walkers into one pipeline.
• Report. orbistoun-shader::report renders the summary and the ranked worklist, in the library so a CLI command and the run report cannot disagree.
• Names. data/mnemonics.toml, generated from what the fixtures observed.

Verified. Gate green. 348 workspace tests. 106 instructions across 9 fixtures, every boundary matching the reference, 14 of 17 encoding families covered, 37 opcodes named. The end-to-end worklist over the fixture corpus renders correctly.

Surprises.

• The contiguity check in the generator earned its place on the first run. The regular expression was anchored to end of line, which silently dropped every branch instruction - they carry a trailing <control+0x2c> symbol reference. The fixture would have encoded a four-byte hole as real and taught the decoder that instructions after a branch start early. A check written on general principle caught a specific bug within minutes of existing.

• LLVM crashes rather than diagnosing when a graphics shader is built for the compute environment. The message "unsupported non-compute shaders with HSA" is there, but it is followed by a stack dump - and the generator was reporting the last stderr line, which is a stack frame. It sent the first investigation straight past the actual answer. Now it prefers a line starting with error:, and the triple is read from each source so a fixture that needs a different environment says so itself.

• The ranked worklist validated D086 on real data without being asked to. s_mov_b32 appears 13 times but in only 2 shaders, and lands ninth. Ranking by raw frequency would have put it near the top and bought one shader's worth of progress. The judgement was made on reasoning; the fixture corpus happened to demonstrate it.

• A mechanical pub const to pub(super) const replacement turned three public methods into dead code. Scoped to a pattern instead of to the private module it was meant for. Third time this session a mechanical transform has overreached - the rule is to scope by meaning, not by text match, and it keeps needing relearning.

• An assertion anchored on the word "shaders" selected the summary line rather than the first blocker, because the summary says "shaders 0 of 9 complete" too. It had a third arm checking for an underscore, which every mnemonic contains, so the test passed while checking nothing. Two weak conditions can hide a broken one.

Not done. No CLI surface - orbistoun-cli/src/main.rs was live in another session. Three encoding families remain unverified. The shader-address register map has no oracle and its output is explicitly a candidate. Nothing is wired into the HLE registry, so no real submission has been walked and no real shader captured.