Documentation audit, first pass

Asked for after a run of changes, and it found the kind of drift that matters rather than typos. Six corrections, all in files that describe the program as it is now:

• README.md and docs/OUTPUT.md both said "nothing in the census is ever called". 910-bulk calls censused symbols by design, and OUTPUT.md documents the record type it emits ten lines below the sentence denying it exists. A format contract contradicting itself on one page is the worst place for this. Both now state the rule and the deliberate exception, including why the exception is a cast rather than a redeclaration.
• README.md status block claimed shadPS4 executes the suite, Kyty relocates and craziiEmu loads - written when that was the whole story. Four loaders now reach REPORT COMPLETE, and one of them no longer loads the corpus module at all (D149).
• README.md census figures were from the 383-symbol era: "all 312 report present", "names the other 238", "several hundred symbols". The census is 39,549 across 372 libraries, and the honest comparison is now shadPS4's 35,337-of-35,337 against PS5PCEM's 31,601 absent - where the higher number is the less honest one.
• CLAUDE.md said "fourteen emulators"; there are eighteen and all are documented. The number was carrying no information and rotting, so it is gone rather than corrected - numbers that inform get gated, numbers that decorate get deleted.
• Two source comments still described the exclusion list as naming checks only, which stopped being true when section entries landed (D145).

Two things the audit did not flag, deliberately. DECISIONS.md and WORKLOG.md are full of counts that no longer hold - 79 checks, 326 censused symbols - and those are dated records whose numbers were true when written; doccheck.py exempts them for that reason and "correcting" them would be falsifying a log. And COMPATIBILITY.md's "499 checks" describes the report it documents rather than the current build, which is also correct.

This is a first pass over the live documents. The per-file source comments are dense and have not been swept end to end.