0.5.2

The last of the audit's corrections, measured on a ninth corpus. A file that documents another project is now said so in one line instead of a wall of findings, and four small families the eighth corpus left behind are filtered.

ANOTHER PROJECT, one line instead of the findings. An auto-detected context file whose cited paths start, for the most part, in folders this repository does not have (at least four of them, and at least six in ten of what it cites) is read as a template waiting to be copied, or a skill written for the codebase it will be installed into. Its findings are held back, a section names the file with the count, and nothing of it reaches the exit code. A file named on the command line, or in targets of .prumorc.json, is always checked in full. The folder is the signal because a note that has gone stale still cites the folders the repository has. The JSON gains elsewhere and schemaVersion is 3.

Small families. "If docs/context.md exists, read it" makes the file's existence a condition and no longer reports it. Double brackets glued to a word, holding a dot or padded with spaces, df[[col]], [[rule.threat]], $[[ inputs.stage ]], are code or template syntax and no longer read as wikilinks. A backtick span that opens with a number, 000 Inbox/Inbox.md, is a name and no longer read as a command. Six more came out of the ninth corpus itself and are marked as unmeasured on the design page: a host written without its scheme, docs.example.com/guide.md, and a file:// address are web addresses; a :symbol after a path points inside it; shot_NN.md is a stencil; "migrated from" and "moved to" are negations; and a quoted argument with spaces no longer splits a make line into a bogus target. For a SKILL.md, its own references/, scripts/, assets/ and templates/ never count toward the gate: a skill missing its own files is a finding.

Measured. Fixed corpora first, 0.5.1 against this release: the fifth and the sixth are unchanged, 16 and 14 findings with the same real ones; the fourth loses only the two hosts read as folders, a family the fourth pass named and nothing had touched since, so the three checks it shares with 0.4.7 stand at 92%; the seventh goes from 2744 to 2536. On the eighth corpus, outside its catalogues, 267 findings to 247, all twenty false, 50% to 54%, and 3807 held back inside them. Then a ninth corpus, sixty repositories from plain searches for SKILL.md, CLAUDE.md and AGENTS.md, none on any earlier list, the gate and the three small rules frozen before it ran: on the forty-four repositories that are not catalogues, 87 findings to 69, real 30 to 29, precision 34% to 42%; in the eight catalogues, 1991 to 1650, and of thirty removals read by hand eighteen were another project's paths, API endpoints written as paths and hosts read as folders, while twelve were a bundle of persona skills linking prompt files it does not carry, which is the trade the gate makes and the design page says so. The three rules 0.5.1 built from the eighth corpus changed nothing on the ninth.

Suite 124 to 131, simulation 102 to 104.