correctness-huntlisted

# Correctness hunt A hunt is not "look for bugs". It is a fixed, repeatable work order. Findings that skip a step do not count. ## The work order 1. **Pull latest and branch.** Fresh branch per hunt (`fix/hunt-<date>` or similar), never the default branch, based on up-to-date `origin/main`. 2. **Baseline before touching anything.** From `vibetags/`: `mvn test -Pe2e`, in the background with output to a log (the run takes minutes and must not be truncated by a tool timeout). Record the counts. A hunt against a red suite cannot separate findings from pre-existing breakage; if the baseline is red, the hunt's first bug is the baseline. 3. **Hunt with named lenses**, not vibes. Each lens is a yes/no question asked of specific code. The standing lens library, from the shapes that have actually bitten: - **Drained evidence**: does any check read state that legitimately drains between the moment of recording and the moment of judging? (The fingerprint/sidecar mtime coupling is this repo's precedent.) - **Re-registration wipes**: does re-running a registration/collection path clear state a sibling still needs? (Issue #365, granular files with multiple authors.) - **Identity from prose**: is anything keyed on a human-readable string that is free to reword, rather than on a structural field? - **One-throwing-participant-discards-all**: does a loop over rules/renderers/modules let one throwing member abort or silently drop the rest?