Normalize code comments and apply safe non-semantic cleanup

A six-phase, constraint-based protocol for normalizing code comments and performing safe, non-semantic cleanup without touching program behavior. Its prime directive is that comments and non-semantic cleanup are the job, logic is never the job - any change that alters what code does, not just what it says or how it's arranged, is out of scope regardless of how obviously "correct" a fix seems. Phase 0 requires reading the entire file or module before editing anything, with no incremental cleanup while still reading, identifying the language's idiomatic doc convention, any existing in-file comment style to match, and any comment encoding real non-obvious information like race conditions or "don't reorder this" warnings. Phase 1 classifies every existing comment into seven categories with a specific action each: junk or venting comments get their real information extracted and rewritten professionally, or deleted if empty; placeholders convert to a proper TODO: or get removed if stale; dead-code comments get removed unless clearly intentional, and are flagged rather than silently deleted; redundant comments get deleted; outdated or wrong comments get rewritten to match current behavior and flagged as stale rather than silently fixed; valuable-but-informal comments have their information preserved with only the tone rewritten; and missing comments get added, without over-commenting self-explanatory lines. Phase 2 limits non-semantic cleanup strictly to changes that cannot alter behavior - consistent indentation, whitespace, and brace style, removing unambiguous dead code, splitting long lines, and local-scope-only variable renaming - explicitly excluding reordering logic, extracting functions, or changing control flow. Phase 3 sets five standards for every comment touched or added: explain why not what, use the language's idiomatic doc format, be concise with no padding, use no informal register, and explicitly avoid AI-tell phrasing like "This function is responsible for..." or em-dashes, plus never inventing a rationale that isn't actually knowable from context. Phase 4 verifies the edited file is behaviorally identical to the original by re-reading the full diff end to end, treating any information dropped from a rewritten comment as a failure requiring restoration. Phase 5 requires reporting back to the user rather than a silent diff: counts of comments rewritten, added, or removed and why, confirmation that informal-but-real warnings were preserved, explicitly listed dead code or stale comments removed, and anything left alone due to uncertainty rather than guessed at.

Use it when the user asks to clean up, professionalize, or polish existing code, code is being prepped for review, handoff, open-sourcing, or documentation, a file mixes human and AI-written comments needing one consistent voice, comments are outdated, missing, redundant, or casual, or the user explicitly wants comments improved without touching logic. Do not use it when the user wants a bug fixed or behavior changed, a full rewrite or architectural restructuring, or only new features added - those are explicitly out of scope. It never renames anything public, exported, or cross-referenced without explicit confirmation, and never deletes a casually-toned comment without first checking whether it carries real information.

Input is an existing source file or module with a mix of comment quality. Output is the same file with comments reclassified, rewritten, or removed per category, plus a narrow set of non-semantic cleanup edits, followed by a summary report rather than a silent diff. For example:

### before
count += 1  # increment count by 1

### after
count += 1

Self-contained - the source states it requires no other skill file to execute the protocol; matches whatever idiomatic doc convention already exists in the target language, such as JSDoc, Python docstrings, Rust ///, or XML doc comments.

Engineers preparing code for review, handoff, or open-sourcing who need comments professionalized and normalized without risking a behavior change slipping in alongside the cleanup.