2026-08-19 - Symbol census: 30 imports to 277

Done. Split the coverage problem in two. The behavioural sections keep asking whether a function works; a new 900-surface section asks only whether it exists.

• 248 censused symbols across 13 libraries, from one X-macro list per library, so a name still appears exactly once.
• Every platform declaration is now weak. An unresolved import is a null address rather than a link failure, and the harness skips a check whose symbol is absent instead of calling it - no more losing every remaining check to one missing function.
• New sym report record, verify.py validates it, census-summary.sh reports the ratio.
• ACKNOWLEDGEMENTS.md written, with the position on names versus signatures versus layouts versus implementation set out explicitly.

Verified. 49 checks, 12 sections, 277 imports (was 30). Report well-formed. Census control passes: 1 present, 247 absent on a host implementing none of the platform.

Surprises.

• The census could not tell "everything is absent" from "the test is broken". The first run reported all 246 symbols absent and every library red. That is the correct answer on this host - and byte-for-byte what a completely non-functional presence test would print. Worse, that ambiguity is at its worst exactly when the tool is most needed: early, when an emulator implements nothing. Fixed with a control that probes one symbol that must resolve and one that must not, defined in a different translation unit so it goes through real weak resolution. (D015)

• Declaring censused names as const char rather than as functions turned a rule into a compiler error. The rule "never call a function whose signature you do not know" was going to be enforced by 248 opportunities to remember it. Declaring them as data means the type system rejects the call outright. The safer design was also the shorter one.

• Two census names collided with real declarations in platform.h - a name cannot be both a function and a const char. Worth having a check for rather than discovering at the next expansion, since the census is expected to keep growing.

• clang-format mangled the census lists, and the damage was not idempotent. It reads X(name) X(name) as chained calls and reflows them into a staircase of continuation indents; formatting and then re-checking still reported violations, so the CI gate could never have passed. Fenced with clang-format off and rewritten one symbol per line. (D016)

• Trying to recover the list by parsing the mangled file lost a third of it. A regex over the wrecked header recovered 213 of 246 names and dropped a whole library

  • and would have passed review, because the file still compiled and the census still produced a plausible number. Regenerated from the authored list instead. The lesson is narrow and worth keeping: when a mechanical transform corrupts a source of truth, reconstruct it from the source, not from the wreckage.

• clang-format over the mount is slow enough to matter - several minutes for 20 files, because every read and write crosses the share. Worth moving builds to a VM-local copy if this grows much further.

Not done. The census is presence only; it says nothing about behaviour, and a platform stubbing every function to return success would score 100%. Struct-taking functions are still absent from the behavioural checks (D008), so the positive-check ratio has not moved. Never run on real hardware or under an emulator.