--version advertises seven frontends the docs never mention

• Track D — prose only.
• Raised by frankD while documenting the information flags ([[docs-toolchain-cli-flags]]), which had to reprint the line. That page carries a one-sentence disclaimer as a stopgap; this ticket is the real answer.

The gap

pxx --version (pinned v391):

frontends:   pascal c nilpy rust zig ada basic fortran algol erlang lolcode whitespace

grep -rn "Fortran\|Algol\|LOLCODE\|Erlang" docs/ returns nothing. The docs cover Pascal, C, Nil Python, and Rust/Zig as experimental. Seven names in that line appear in no public page.

They are not vapour — each was fed a source file with its own extension and answered with a frontend-specific diagnostic (Fortran: unexpected statement, Erlang: no main/0 found, Algol's own parser error), so the lexers and parsers really run. But the reader has no way to tell what any of them is for, and the list flattens two very different things:

Why it is worth deciding rather than ignoring

The string is hand-maintained — compiler/compiler.pas:272, a literal, not a registry read — so it is already a drift risk, and it is the most public surface the esoteric frontends have. Left as is, --version is the only place a user learns these exist, and it reads as a support claim for Ada and Fortran.

Options

1. A short "other frontends" page under docs/targets/, saying plainly which are demo targets and which are probes, and that neither carries a compatibility promise. Cheapest honest fix; the tickets already contain the material.
2. One paragraph in docs/targets/index.md, no new page. Less discoverable, but the list is short.
3. Leave the docs alone and change what --version prints — group the line, e.g. frontends: pascal c nilpy (rust zig experimental; +7 probes). That is Track A's file, so it would be filed there; recorded here because it is the alternative that actually removes the gap rather than documenting around it.

Recommendation: option 1, and it is small — a page that exists mostly to say "these are probes, do not build on them" is worth more than the silence, which currently reads as an omission rather than a decision. Option 3 is a good companion but not a substitute: the names stay reachable either way.

Not in scope

Whether any probe should be promoted, and anything about the frontends' behaviour. This is about what the public docs say exists.

Log

• 2026-08-29 — resolved, commit 05dbd4c0e.

RESOLVED 2026-08-29 (frankD)

Option 1 as recommended and as the coordinator confirmed: a short docs/targets/other-frontends.md (order 66), linked from docs/targets/index.md and from the --version paragraph in docs/reference/cli.md. Option 3 — changing what --version prints — was not pursued: that is compiler/compiler.pas:272, Track A's file.

What the page says

A four-row table separating the kinds, then three short sections. The line it had to draw is not "documented vs not" but real frontend vs skeleton probe, and the evidence for it was already in the repo:

So BASIC gets its own section with a worked example, and the six get a shared one.

The framing that makes the page worth having

Taken from the umbrella (feature-esoteric-frontend-probes), because a list of curiosities is not worth a page but this is: the point is not that PXX compiles Fortran. The probes exist to prove the shared AST and IR are correct by feeding them programs shaped unlike anything the mainline frontends emit — column-sensitive lexing, implicit typing by first letter, dynamic loose casting, GOTO-first control flow. "Oh, and it compiles Fortran" is the side effect; the differential test of the pipeline is the deliverable. That is the same argument nil-python.md already makes for itself, so the page points at it.

The honest summary is stated in the page rather than implied: it works, it is one test wide, and it is not going anywhere — plus "if you want to compile Ada, use an Ada compiler", which is the sentence a reader needs and the one a list alone never says.

Measured on v393 — nothing described here is second-hand

• all six skeletons compile and run: sum correct (ada), 55 (algol), fact(5) is 120 (erlang), sum is55 (fortran), HAI WORLD (lolcode), Hi (whitespace);
• BASIC: test_basic_comprehensive.bas and test_basic_goto_gosub.bas both compile and run;
• the page's own BASIC snippet was extracted from the rendered Markdown, compiled and run — prints A then looped 3, with the SKIPPED line correctly jumped over;
• the quoted --version line was string-compared against real output. It did not match on the first pass — I had dropped the two leading spaces while presenting it as program output. Fixed and re-checked; the full --version block quoted in cli.md was verified the same way at the same time.

One correction to this ticket's own framing

The ticket's table guessed that ada, erlang and whitespace were "not checked when this was filed". They are all three esoteric probes (feature-esoteric-ada, feature-esoteric-whitespace, feature-erlang-frontend-scoping — the last records the skeleton as done and full-language scoping as untouched). And BASIC is stronger than the ticket claimed: the umbrella names Pascal, C, BASIC and Nil-Python as the first-class test surface, so BASIC is core, not merely "a real demo target".

The umbrella also lists Zig among the opportunistic set, which is stale — Zig and Rust have had their own tracks since. The page follows CLAUDE.md's current structure (R and Z experimental) rather than the 2026-07-05 snapshot, and says their difference from the probes is ambition rather than current support.