raeq

epubveri 0.16.0: four changes that cut CPU time to 39%

epubveri-0.16.0-perf.patch applies to the v0.16.0 tag with git apply or patch -p1. It changes no finding: over a 2,798-book Apple Books library, every report matches 0.16.0's (message, severity, file and line) on every book. The crate's own tests pass, 711 of 711, and clippy reports no new warning.

Where the time went

The slowest book on the shelf, Python (521 XHTML files, 108 s of CPU), sampled with macOS

raeq

epubcheck message ids: 21 of 315 use an underscore

repro.sh builds a minimal fixed-layout EPUB with two faults and runs epubcheck 5.3.0 on it. Output:

ERROR    HTM_056    Viewport metadata has no "height" dimension (both "width" and "height"
ERROR    RSC-012    Fragment identifier is not defined.

raeq

"""weaponizing-unicode against the normative confusables.txt: what the corpus is, and what each detector covers.

Reads the paper's `new_predicted_homoglyphs.txt` from $DISARM_META_WEAPONIZING or the meta-benchmark
cache, and UTS #39 `confusables.txt` from $DISARM_META_CONFUSABLES or `data/confusables.txt` in the
checkout. Needs disarm and confusable-homoglyphs importable. Expected output on disarm 0.15.0 at
fba73bc5 with confusables 17.0.0 and confusable-homoglyphs 3.3.1:

    scored 7402; UTS #39 sources 586; single-code-point prototype 39; of those in another script 35
    multi-code-point-prototype sources by script: [('ARABIC', 193), ('CANADIAN', 129), ('HANGUL', 108), ('LATIN', 20)]
      scored                         n=7402  disarm  236  confusable-homoglyphs  362

raeq

"""Every disarm surface on the Emoji Attack construction: an Emoji_Presentation code point inside a word.

The ladder is the meta-benchmark's (`benchmarks/meta/suites/academic.py`, EmojiDelimiterSegmentation):
rejoined (emoji gone, word restored) > survives (still two runs) > widened (a name or gloss:
extra runs) / substituted (became letters) / destroyed (carrier lost). Reads emoji-data.txt from
$DISARM_META_EMOJI_DATA, else downloads it from unicode.org. Needs disarm importable; run from a
checkout for the profile names. Expected output on disarm 0.15.0 at 43e24b6b:

    disarm 0.15.0; Emoji_Presentation code points: 1219
    surface                               rejoined  survives  widened  subst destroyed  aa🔥bb ->

raeq

"""Re-score `mcp-tag-block-concealment` under three parcel rules, with the harness's own code.

Run from a disarm checkout that has `benchmarks/meta` (main at 2b2db6cf or later). The
measurements are quoted from the 2026-09-04 run of `python -m benchmarks.meta --run
--subject all` on disarm 0.15.0 at 2b2db6cf, so no run is needed. Expected output:

    A. as scored today (detected_t1_plain: flagged = good)
      one parcel             disarm rank 8 of 12 (parcel +0.026); top +0.760; confusable-homoglyphs rank 7 (+0.707)
    B. detected_t1_plain inverted (not flagged = good)
      one parcel             disarm rank 7 of 12 (parcel +0.733); top +0.760; confusable-homoglyphs rank 8 (-0.707)

raeq

"""benchmarks/meta composes three disarm pipelines with strip_zalgo=0, which strips every diacritic.

disarm documents max_marks=0 as "strip all combining marks" (docs/api/transforms.md:
strip_zalgo("café", max_marks=0) == "cafe"), and None as off. The composed subjects in
benchmarks/meta/subjects.py pass 0. Run from a disarm checkout that has benchmarks/meta,
with disarm importable. Expected output on disarm 0.15.0 at da174c83:

    disarm 0.15.0
    TextPipeline(strip_zalgo=None) steps=[]                           'Čeština, naïve café'
    TextPipeline(strip_zalgo=0   ) steps=[('strip_zalgo', '0')]       'Cestina, naive cafe'

raeq

"""is_confusable fires on printable ASCII: `"`, `|` and backtick are TR39 sources.

Run from a disarm checkout (the four prose files are read relative to it) with disarm
importable. Expected output on disarm 0.15.0 at da174c83 (tables Unicode 17.0.0,
confusables 17.0.0):

    disarm 0.15.0
    is_confusable('"')                                 True
    is_confusable('|')                                 True
    is_confusable('`')                                 True

raeq

#!/usr/bin/env python3
"""disarm #937: Boucher et al.'s four classes scored one at a time, and a cursor model
for the deletion class.

The meta-benchmark scores the released corpus as one population. Split by the
experiment key, the aggregate is the sum of two classes near 100%, one near 0%, and
one at 0% — and the two near zero are the two whose *rendering* is the clean string.

Corpus: https://raw.githubusercontent.com/nickboucher/imperceptible/main/results/adversarial-examples.json
(MIT). Nested experiment -> budget -> row -> {"adv_example", "input", ...}. Pass the path

raeq

#!/usr/bin/env python3
"""#739's detection half shipped in #934; the recovery half did not (revised 2026-09-03).

Boucher et al.'s deletion class hides a character from a human reader by
following it with an erasing control. disarm now *detects* every form of it and
*names* the semantics — "erases the preceding 'a'" — and no surface applies
them, so the smuggled character reaches whatever reads the output.

    pip install disarm==0.15.0
    python deletion_recovery.py

raeq

#!/usr/bin/env python3
"""Reproduce every number on the disarm meta-benchmark coverage/cost plot.

The plot shows two of thirteen axes. This script recomputes both of them, the
composite, the bootstrap intervals and the Pareto frontier from a run of the
harness, and prints them in the order the page's table lists them.

    git clone https://github.com/raeq/disarm && cd disarm
    git checkout bench/meta-harness
    python -m venv .venv && .venv/bin/pip install -e . maturin