max / makenotwork

Fuzz tagtree: an oracle, two targets, and the seeds to replay them on stable Row 8 of astra-soak-overview. The witchbroom side landed in witchbroom@560c058; this is the harness it schedules. THE DOORS WERE COUNTED FIRST, as the task requires and as git_ssh taught. 12 files across 5 repos call tagtree -- MNW server and multithreaded, audiofiles, balanced_breakfast, goingson -- and every one reaches the same implementation. Nothing needed extracting, so one target covers the whole string surface. What the count did change is the input: five distinct TagConfig constants are live in the tree, so oracle::CONFIGS enumerates all five plus validate's default and four degenerate ends, and every input is checked against all ten. Two targets, because the input shape differs and not because the grammar does. A tag string and an operation sequence over a mutable TagIndex cannot share a corpus without most of its bytes landing on the wrong half. Both take plain text rather than a derived Arbitrary struct, so the seed corpus is the 130 production slugs from MNW's migrations as files a reviewer can read in a diff. THE ORACLE FIRED ON CORRECT BEHAVIOUR ON ITS FIRST REPLAY, which is the useful part. common_ancestor("a.b.c", "a.b.c") returns Some("a.b"), not None: the byte scan runs to the end of both strings, so the last separator it recorded is the one before the final segment. The model had generalised from the crate's documented depth-1 example, and it was corrected to the crate's behaviour rather than exempted from it. Silence was then confirmed the hard way -- all 16,900 ordered pairs of the 130 slugs, and the whole taxonomy queried at every prefix at three limits -- and both are committed tests. No bug was found in tagtree, so regressions/ ships with its README and nothing else. The teeth are three full reimplementations rather than invariants. model_suggest filters the whole tag list where the crate binary searches and breaks early; model_common_ancestor counts segments where the crate scans bytes; and escape_like is checked by un-escaping under its own documented rule, which is the property the seven live SQL call sites lean on and the one reached by unvalidated user text in audiofiles' free-text search. model_suggest_fuzzy scores with a naive full-matrix Levenshtein against edit_distance's single-row DP, aimed at the mutants that survive there. It works on BYTES because edit_distance calls as_bytes; a chars() reference would disagree on every non-ASCII input and report correct behaviour as a finding. One invariant is deliberately weaker than it looks: the segment index is a SUPERSET of the segments of the live tags, never an equality, because remove documents that it does not prune orphaned segments. An equality assertion would fire on the first removal and every report after it would be noise. Also corrects docs/architecture.md, which had the semantic_depth column wrong in both directions -- AF listed 1 and is 0, MNW listed 0 and is 2. Verified against the constants themselves. The 2 is what forces the three-segment type.category.value slug shape on the platform, which the oracle had to know to enumerate the configs at all.