Non-ASCII ban: whitelist box-drawing chars and currency symbols

[MPZ-00]

@eugeniughelbur - touches CLAUDE.md so wanted your take first.

The em-dash ban does not stick. Claude slips them in constantly, I end up grepping after the fact. A hard 7-bit ban is cleaner to enforce.

Problem is that 7-bit blocks things I actually use: ─ (U+2500, alt+196) for separators in structured text, and € or £ when noting prices. Not substitution Unicode, they mean something concrete.

My setup-dotfiles/CLAUDE.md draws the line like this: em-dashes, en-dashes, curly quotes, smart apostrophes, and Unicode math fakes (>= not >=) banned everywhere. Box-drawing chars and Nerd Font codepoints fine in terminal configs.

Worth aligning on this for the repo? Happy to PR if yes.

[eugeniughelbur]

Yes, ship it. Your read is right and I have the receipts: I just checked our current vault notes from work earlier today and counted 28 em-dashes across two files I wrote myself, after explicitly trying to avoid them. The rule does not stick. Mechanical enforcement is the only path that works.

Your whitelist is honest. Box-drawing chars and currency symbols have semantic meaning, not LLM-substitution Unicode. Drawing the line there beats either extreme.

PR structure I would prefer:

1. First PR — the rule + validator. Extend hooks/validate-ai-first.sh with a non-ASCII check that respects your whitelist (box-drawing range U+2500–U+257F, currency block, and whatever Nerd Font ranges you want explicitly named). Update CLAUDE.md to swap the "no em-dashes" line for the explicit whitelist rule. Update references/ai-first-rules.md similarly. Keep this PR tight — no codebase sweep yet.

2. Second PR — the sweep. Run the validator (or a sed-based sweep) across every tracked *.md / *.py / *.sh in the repo. Convert — → -, smart quotes → straight, ≥ → >=, → → ->, etc. Manual pass over anything ambiguous. The two-PR split means a bad substitution in the sweep cannot block the rule from landing.

Splitting like that also means the sweep PR can be safely reverted if anything reads weirdly without losing the rule itself.

One thing I want flagged in the validator output: the specific codepoint and a one-line "did you mean" suggestion, so when someone trips it they know whether they violated by accident or intent. U+2014 em-dash → suggest -. U+201C left double quote → suggest ". That kind of thing.

Open both whenever you are ready.

— Eugeniu

[eugeniughelbur]

Merged. PR #43 is on main as commit da33d64. Thanks for the fast turnaround - exact shape I wanted, codepoint output works, whitelist holds.

Standing by for the sweep PR whenever you are ready. 852 violations across 90 files is a real budget but mostly mechanical - the per-file overhead is small, and your "fork-priority on conflicts" approach already maps to "fork-style decisions on borderline substitutions." If anything reads weird after a substitution, fix it manually in the same PR rather than reverting; one bad substitution should not block the whole sweep.

One small nit you might roll into the sweep PR or a separate cleanup, no obligation: the allowed(cp) function in check 5 is documentation, not enforcement - any character not in the BANNED dict already passes via the ch not in BANNED clause. Behavior is correct but the function never gets reached in practice. Could either be removed as dead code or the design reframed as "explicit ban list, anything else allowed" rather than "whitelist + blacklist." Up to you; not blocking anything.