# Trivia-invariant-layout gate baselines (S0)
The failure inventory the S1–S4 stages of the trivia-invariant-layout plan
(TODO.md) are gated against: **compare sets, not counts** — a stage may shrink
these sets but must not grow them.
Recorded at badness commit `268c5d8` (S0); re-recorded after S2, which removed
`latexrelease.sty` from both latex3 sets (15 `all`, 150 `trivia`) and refreshed
`SWEEP.md`. Re-recorded again after S4: the `all` sets are byte-identical to
S2's, and the `trivia` sets trade four statement-boundary entries resolved
(`xtemplate-2023-10-10.sty` in both corpora — the S4 motivating case — plus
latex2e's `pdfmanagement.sty` and `tagpdf-base.sty`) for four additions in a
different, pre-existing family (`xparse-2020-10-01.sty` in both corpora,
latex2e's `lipsum.sty` and `expl3.sty`): the out-of-region prefix mode/rest
printer coupling filed as an S4 follow-up in TODO.md, reachable only through the
all-newlines-to-spaces mega-line variants. Re-recorded once more after the
grouped-sibling-walk fix (`head_command_has_grouped_sibling_arg` now segments
the same stream the layout segments — container braces stripped, sliced to the
in-region run): three `non-fixed-point` entries resolved
(`xparse-2020-10-01.sty` in both corpora — two of the four S4 additions were
this, not the mode/rest coupling — plus latex2e's long-standing `xparse.sty`),
no additions, and production output is byte-identical over all three corpora at
widths 60/80/120. Over the pinned gate corpora fetched by
`task gate-corpora:fetch` (`scripts/fetch_gate_corpora.sh`):
| latex3 | `latex3/latex3` @ `3d1d347d8937863c0786988b14d307a6091ee397` | 288 |
| latex2e | `latex3/latex2e` @ `3a9fdd88bdc53f16a0c2158aa70d259607de333a` | 384 |
| pgf | `pgf-tikz/pgf` @ `1c7fc0fdc3ec8a6bdcfd68785c6bbd43ec110178` | 397 |
## Checking and regeneration
The ratchet is machine-checked: `task gate-corpora:check`
(`scripts/check_gate_baselines.sh`) re-runs both gates over the pinned corpora
and diffs the distilled failure sets against these files, two-sided — an added
line is a regression, a removed line is a stale baseline, and either fails the
check. It needs the corpora (`task gate-corpora:fetch`) and network, so it is a
pre-merge/re-record step for formatter changes, not part of `cargo test`.
To regenerate by hand, from each corpus root with a release build:
```sh
badness --no-config debug format --checks all --report . # -> <corpus>.all.txt
badness --no-config debug format --checks trivia --report . # -> <corpus>.trivia.txt
```
Distill each report's `### k. \`path\`
(kind)` headings into sorted `path<TAB>kind` lines (`.all.txt`), adding a third class column for the trivia files from each failure's `-
Variant:` reason (`.trivia.txt`) — the same distillation `check_gate_baselines.sh` implements, so its diff output can be applied to the baseline files directly. The width sweep behind `SWEEP.md` reruns both checks with `--line-width
60\|72\|100\|120\`.
## The sets
- `<corpus>.all.txt` — the pre-existing `--checks all` gate (losslessness +
idempotency at width 80). At S0, latex3's 16 (15 `format-error` +
`latexrelease.sty` idempotency) matched the baseline recorded in TODO.md
before S0, i.e. S0 changed no production layout; S2 resolved the
`latexrelease.sty` entry, leaving latex3 at 15 `format-error`.
- `<corpus>.trivia.txt` — the `--checks trivia` convergence-oracle inventory at
width 80 (wrap pinned to reflow). Counts: latex3 149 (151 at S0), latex2e 146
(148 after S4), pgf 15.
- `SWEEP.md` — failures that appear or vanish across widths 60–120; each is a
column-arithmetic hybrid candidate.
## Classification (third column of `.trivia.txt`)
- `format-error` — the formatter refuses the file (statically unmodelable
constructs; matches the smoke-test workflow's ALLOWLIST families). latex3 15,
latex2e 13, pgf 15. pgf has **no** trivia failures at all.
- `content-change` — `fmt` violated the whitespace-only invariant on a perturbed
input (latex3 132, latex2e 117). Two root causes identified, both
**pre-existing reflow-on-`.dtx` doc-layer bugs**, reachable without any
perturbation via `badness format --wrap reflow`:
1. prose reflow relocates `^^A` doc comments into positions where they re-lex
as content (the l3 `\title{...^^A...}` preamble pattern — essentially every
l3 `.dtx`);
2. content on docstrip-guarded lines is dropped entirely (`alltt.dtx` loses
`\ProvidesFile{alltt.drv}`). These dominate the trivia sets and mask any
later failure in the same file (the check stops at the first failing
variant), so fixing them will both shrink these sets and potentially
surface currently-hidden non-fixed-points.
- `non-fixed-point` — a perturbed variant formats to a non-fixed-point: the
idempotency-hybrid family the umbrella exists to fix (latex3 3 after S2,
latex2e 18; predominantly expl3 package code — xparse, xtemplate, tagpdf,
pdfmanagement, luamml — the S2–S4 target set). `latexrelease.sty` was the
known issue-#97 residue (rest-aware fill disagreement); S2 resolved it.
The generator's meaning-safety was spot-checked: 1 `dropped_unsafe` variant in
\~157k eligible gaps over 120 latex2e files.