mastodon-typed-text.lua

Reconstructs what an author typed into the Mastodon compose box from the
rendered HTML stored in an ActivityPub archive, and counts that text the way
the compose box counted it while the author watched. The archive stores what
the server rendered, not what was typed: markdown emphasis delimiters are
consumed into tags (*love* becomes <em>love</em>), URLs are wrapped in
anchor markup with invisible spans, and entities are escaped. Recovering the
typed text is what makes golden poem qualification (exactly 1024 characters as
composed) possible. Created for issue 4-003 (August 2026 re-open).

External functions

restore(html) → string

• Input: html (string) — one post's content HTML as stored in the

archive's outbox.

• Output: string — the reconstructed typed text: emphasis delimiters put

back (*, **, ~~, backticks), paragraph breaks as blank lines, line
breaks as newlines, entities decoded (& decoded last so a typed
< survives), all remaining markup stripped, leading/trailing newlines
trimmed. Typed backslashes are preserved (the older inline cleaner deleted
them before quotes).

• Used for both the displayed poem content and the golden poem content.

compose_box_count(html, content_warning) → number

• Input: html (string) — post content HTML, pre-anonymization so real

mention text is priced; content_warning (string or nil) — CW text without
any "CW: " prefix.

• Output: number — the count the author watched while composing:

reconstructed typed text plus CW text, measured in composer characters,
with every non-mention link priced at a flat 23 regardless of URL length,
and mentions priced at @user alone — both anchor-form mentions (visible
text is already @user) and plain-text @user@domain mentions the server
never linkified (the domain part rides free).

• A poem is golden when this returns exactly 1024.

composer_length(s) → number

• Input: s (string) — UTF-8 text.
• Output: number — length in composer characters, the unit the compose

box counter used: one per visible character regardless of byte width
(a curly apostrophe is 3 bytes but counts 1), one per emoji even from the
astral plane (4 bytes, counts 1), zero for the invisible glue codepoints
emoji carry (variation selectors, zero-width joiner). Empirically anchored:
archived poems with emoji sit at exactly 1024 under this model and at 1025
under UTF-16-unit counting, and the compose box refused anything past 1024.

Data notes

• Mention anchors carry class="u-url mention"; hashtag anchors carry

class="mention hashtag". Both keep their visible text; only true URL
anchors become the 23-character placeholder.

• The legacy ^_^ emoticon repairs from the original cleaner are retained

verbatim (observed archive-specific damage; length-neutral).

• Underscore-style emphasis (_x_) is indistinguishable from asterisk-style

in rendered HTML; both restore as asterisks — identical length, so counting
is unaffected either way.

Tests

libs/mastodon-typed-text-test.lua — standalone, plus an optional
integration pass that recounts three audit-verified poems straight from
input/fediverse/files/poems.json when that file is present.