2026-08-19 - Test corpus and the accuracy suite

Documentation only; no code touched.

Done. D042-D045 recorded. Corpus tooling (titles/, manifest-driven, pinned by hash) and the accuracy suite decided as a separate repository, obSCEne, with its result model and coverage strategy specified. Roadmap gained phases 0d (corpus tooling) and 1b (corpus-wide survey report). titles/ added to .gitignore and to the provenance guard's exemptions in all three places it is enforced.

Surprises.

• No accuracy suite exists for this target. Checked rather than assumed - nothing comparable to pspautotests, blargg's ROMs, or dolphin-emu/hwtests. shadPS4 has host-side unit tests, which is a different thing entirely. That is why we are building one.
• A test app we compile ourselves is a real container with clean provenance, which partly dissolves phase 0's problem: genuine vendor-format input for the parser without anything of anyone else's in the repo. Synthetic fixtures are still needed for malformed cases a compiler will never emit.
• The open toolchain is a lawful source for interface facts - names and arities - because it is open source. That converts two "someday" items (D025's name list, the 37 provisional arities) into tractable ones. The line is D044: interface facts yes, implementations never.
• The reporting channel is itself under test. A suite reporting via stdout says nothing when stdout is unimplemented. Resolved by trace-as-report: orbistoun already records every call, so calling interfaces in a known order is the result, with beacon calls (sentinel argument values) as trace delimiters. Works with zero I/O.

Next. Unchanged - phase 0, with 0b, 0c, and 0d available in parallel. obSCEne awaits scaffolding.