nx_toolhealth.nx

nx_toolhealth.nx -- FOUNDATION CENSUS over the REGISTERED-TOOL POPULATION (lane D, 2026-08-20). The operator named three symptoms in one sentence -- "broken tools and duplicate tools and things like the research only are half capable" -- and the estate counted them in separate censuses that never adjudicated as ONE foundation. This organ measures the two that had no instrument at all, over EVERY row of tool_allowlist.conf (the GREEN execution allowlist == the registered-tool registry), dedupes REFERENCES to SUBJECTS, and prints a partition that must SUM. BROKEN, per SUBJECT, one class each and the four sum to the subject count: PHANTOM no elf anywhere AND no .nx -- the CALLER must be fixed or deleted; it can never be built, and hiding it in a build queue guarantees the queue never finishes. MISSING-BUILDABLE no elf, source exists -- a build. BINARY-NO-SOURCE elf present, no .nx -- often code that SHIPS INSIDE ANOTHER ARTIFACT. Its own bucket on purpose: a stat-based catalogue reads this as "build it", which is a harmful work order, so it is NEVER folded into MISSING-BUILDABLE. OK elf and source both present. UNREACHABLE is a SEPARATE AXIS, never a partition member -- a subject can be OK and non-GREEN at once, so folding it in would silently break the reconciliation. COUNT ONLY (lane A owns the repair). HALF-CAPABLE: a verb DECLARED in the SERVED call grammar with no dispatch branch in the source. The served grammar is resolved exactly as nx_tools_api resolves it -- authored tool_schemas.conf col8 first, derived tool_grammar.conf second -- so this measures the contract an AGENT actually reads. HOW A VERB IS JUDGED IMPLEMENTED: its name appears in the source as a QUOTED STRING LITERAL (the byte sequence quote-verb-quote). That is what every dispatch in this estate compiles to, and a usage line -- which is a COMMENT, or a single long "usage: a | b" literal -- can never satisfy it. THE DETECTOR IS DELIBERATELY WRONG IN THE DIRECTION OF NOT ACCUSING. Four abstentions, each its own bucket, because a false positive teaches everyone to ignore the detector: PINNED the allowlist row carries a 4th column, so the registry supplies argv and the caller's is DISCARDED. Its grammar names no callable verb; it can have no dead verb by construction. NO-GRAMMAR neither grammar file has a row -- an agent has NO contract to read. A real gap, listed. POSITIONAL-ONLY a contract exists and declares no VERB. NOT a defect -- and NOT the same bucket as NO-GRAMMAR, whose remedy is the opposite. Merging them read 618 of 959. (this bullet is indented FOUR spaces on purpose: at 2-3 spaces the grammar harvester reads a comment carrying < or [ as a USAGE LINE, so this very explanation became this organ's declared call grammar and the census then accused ITSELF of a dead verb named POSITIONAL-ONLY. A TOOL WHOSE INPUT IS ITS OWN DOCUMENTATION MUST OBEY THE DOCUMENT CONVENTION IT MEASURES -- caught by the ratchet, not by review.) UNDECIDABLE the grammar declares 2+ verbs and NOT ONE of them appears as a literal => this organ cannot see this subject's dispatch mechanism at all. Reporting N dead verbs there would be a confident wrong answer; abstaining names it for a human instead. RATCHET: the baseline is a SET OF NAMES, never a count -- a count-only ratchet on a shared tree reports a regression without saying whose. Removals ratchet down automatically; ADDITIONS keep firing until fixed or explicitly adopted with `accept`.