alkblobs/tasks/unicode-hygiene.md

id, name, status, depends_on, scope, risk, impact, level

id	name	status	depends_on	scope	risk	impact	level
unicode-hygiene	Gitea-flagged confusable unicode out of the repo prose	completed		narrow	trivial	project	implementation

Description

Gitea's file-view banner ("This file contains ambiguous Unicode characters") flags characters via its confusable tables (modules/charset/generate/ambiguous.json). Checked against that table at decomposition (2026-10-16), exactly three characters present in this repo's prose are flagged; everything else in the house typography is not:

• U+2013 en-dash — confusable with -. 16 in src/, 4 in tests/, 78 in docs/, 32 in tasks/ (the flagged pair the user saw: src/dispatch.rs lines 9 and 28, both in "~128-256 KiB").
• U+00D7 multiply sign — confusable with x. 6 in src/, 75 in docs/, 34 in tasks/ (e.g. "20-50x behind page-cache fs").
• U+2228 logical or — confusable with v. 2 in src/ (src/gc.rs 351-352, "(pins or markers or sources)"), 4 in tasks/.

Not flagged and therefore untouched: em-dash (U+2014, the house prose style), §, →/⇒/↔, …, ≤/≥/≈/≠, box-drawing, ±, µ, ·, and the accented letters in quoted names. This is a surgical replacement, not a typography overhaul.

Replacements, meaning-preserving:

• en-dash → ASCII hyphen (ranges read "128-256 KiB").
• multiply sign → x ("20-50x behind").
• logical-or → the prose reading ("or") where it reads better — src/gc.rs's "(pins or markers or sources)" becomes an "or"-shaped clause; the exact rendering is the task's design point, recorded in Notes.

Scope: all tracked text files — src/, tests/, docs/, tasks/, fuzz/'s tracked text files (README, scripts, manifests). Never touch fuzz/corpus/ (binary seeds) or anything under fuzz/target/ (gitignored build artifacts).

Record the rule in AGENTS.md (one line near the conventions: the three gitea-flagged confusable characters stay out of repo prose; ASCII equivalents; em-dash/§/arrows remain house style) so new prose stays clean after this task.

Acceptance Criteria

• grep -rP '[\x{2013}\x{00D7}\x{2228}]' over tracked text files (src/ tests/ docs/ tasks/ fuzz/ — text files only) returns nothing
• em-dash / § / arrow / ellipsis typography untouched (the house style survives; spot-check src/store/mod.rs and docs/architecture/overview.md)
• AGENTS.md carries the three-character rule
• Full verification block green (cargo test, both clippy variants, cargo fmt --check) — the diff is comments/prose only

References

• gitea modules/charset/ambiguous.go + generate/ambiguous.json (the confusable tables the banner reads)
• src/dispatch.rs (the reported lines 9/28), src/gc.rs (the logical-or pair), src/pg_lo/mod.rs / src/postgres/mod.rs (the multiply-sign cluster)

Notes

To be filled by implementation agent

• Replacement counts came in close to the decomposition estimate (en-dash 30 lines, multiply sign 36 lines, logical or 3 lines by git grep -c — line counts, not occurrence counts; the multi-hit lines are the dispatch.rs/gc.rs pairs the user flagged).
• Rendering decisions: en-dash and multiply sign were mechanical (- and x). For the logical-or pair in src/gc.rs 351-352 and its two prose mirrors (tasks/gc-protocol.md, tasks/review-wave-2.md); the "(pins or markers or sources)" clause was rendered as "(pins / markers / sources)" — the slash-list is an existing house idiom (local/mem backticks) and reads better than a doubled "or" in a parenthetical. In tasks/review-wave-2.md the sibling "(pins / delete markers ∞ sources ...)" keeps its interposed "infinity" glyph untouched — only the three flagged characters were in scope.
• fuzz/'s tracked text files (README, run-detached.sh, gen_fuzz_seeds.py, manifests) had zero occurrences; the corpus and gitignored build areas were never opened for edit.
• tasks/unicode-hygiene.md itself carried occurrences (its own description quotes the flagged characters); those mentions were rewritten to codepoint/ASCII phrasing so the acceptance grep stays clean including this file. The AGENTS.md count line ("pg-heavy tests x up to 2 pools each x pool max_size 8") keeps the arithmetic reading.

Summary

To be filled on completion

Completed 2026-10-16. All three gitea-flagged confusable characters (U+2013 en-dash, U+00D7 multiply sign, U+2228 logical or) removed from the repo's tracked text files: mechanical ASCII replacements (-, x) everywhere, plus the slash-list rendering for the src/gc.rs logical-or clause and its prose mirrors. Diff is comments and prose only — no code changes. The three-character rule recorded in AGENTS.md convention 11. Acceptance grep (git grep -nP '\x{2013}|\x{00D7}|\x{2228}') returns nothing across src/, tests/, docs/, tasks/, fuzz/; em-dash/§/arrow/ellipsis typography spot-checked intact in src/store/mod.rs and docs/architecture/overview.md. Full verification block green.