Comment generation does work it discards, and degrades silently in a shallow clone

[mairas]

From the review on PR #1: #1 (comment)

Measured on synthetic repositories at halpi2's shape and a plausible larger one.

• english_diff was memoised in PR Package the documentation checkers and add a blocking translation gate #1 on (stamp, page) — nine locales stamped against the same blob had each paid a git cat-file, a git diff --no-index and a temp directory for identical output. 200 pages × 9 locales went from 44.94 s to 6.32 s. Done; recorded here because the remaining items share its cause.
• --comment computes every diff before learning none will fit. collect(want_diff=True) does all the git work up front and render_comment then discards it. 20 × 9 all stale: 3.3 s and 360 subprocesses spent on diffs, zero diff blocks in the output. Render the table first, measure it, and compute diffs only within the remaining budget.
• Every diff is buffered in memory at once. One 6.9 MB page fully rewritten across 9 locales peaked at 483 MB RSS for a 1741-byte body; two such pages, 874 MB. Cap the retained diff per entry — nobody reads a 14 MB diff inside a collapsed <details>.
• A depth-1 checkout silently loses every diff. actions/checkout defaults to fetch-depth: 1. The verdict survives, because hash-object reads the working tree, but every stamped blob lives only in history, so git cat-file fails on all of them and the report carries no diffs. The only notice is an HTML comment, invisible in rendered markdown, and neither README nor AGENTS.md mentions fetch-depth anywhere. Emit a visible line and document the requirement.
• check-anchors reads every built page into memory twice. Fine at 36 pages plus an export; the same unbounded-buffer shape as the diff path.