meta-consolidate-claude-mdlisted

# Consolidating duplicated CLAUDE.md content Copy-pasted guidance does not stay identical. It forks, and each fork ages separately, so the duplication problem becomes a CORRECTNESS problem: some copies now describe a tool that changed, a file that was deleted, a target that never existed. Deduplicating without checking which copy is right just picks one lie and installs it at an ancestor, where it binds every repo below. **The order is: measure, verify, converge, lift.** Skipping straight to the lift is the failure this skill exists to prevent. ## 1. Measure - byte-identical groups and their true common ancestor `scripts/claudemd_variance.py` does this measurement, so it never needs hand-rolling again - two sessions wrote this exact script from scratch before it shipped: ``` uv run scripts/claudemd_variance.py --root ~/src --json ``` It splits every `CLAUDE.md` into `## ` sections, hashes each body (whitespace-normalised, so trivial reflowing does not read as a different variant - the definition is in `--help`), groups the identical ones, and reports each group's common ancestor plus the largest variant's share of the group. For each group of 3+, its common ancestor - not the one you assumed - is where the text belongs; `--lift-threshold` marks which variants clear that bar. Enumeration is a plain filesystem WALK, never the session `grep` or any gitignore-aware tool: `grep` routes to a gitignore-aware backend and silently drops ignored files. Measured on one tree: 73 f