A bad value for a known option is reported as unknown option

• Type: bug — Track A (compiler/compiler.pas, argument parsing).
• Found 2026-08-30 by frankD while sweeping docs/** for documented flags the compiler rejects ([[bug-d-the-cli-reference-documents-a-flag-the-compiler-rejects]]). Measured against $(PXX_STABLE); no rebuild.

Measured

The last row is the problem: a real diagnostic and a wrong one are the same sentence. --esp-profile shows the right behaviour exists — it validates its value and says something specific — so this is an inconsistency between option handlers, not a missing capability.

Why it costs more than a wording nit

• It aims the reader at the wrong half of the argument. unknown option means "this flag does not exist", so the reader checks spelling, greps docs, and concludes the documentation is stale — when the flag is right and the value is wrong. --target= is the flag a new cross-compiling user types first.
• It makes the CLI unprobeable. Anything enumerating accepted options by running them — a doc check, a completion script, an agent auditing --help against reality — cannot distinguish a nonexistent option from a live one given a placeholder. This is not hypothetical: it produced a false finding in the sweep that found it. --xtensa-cpu=lx6 is documented correctly at docs/reference/cli.md:122, was probed as --xtensa-cpu=esp32, answered unknown option, and was one step from being filed as a documentation bug. The wrong answer was plausible, which is this repo's expensive class.
• It interacts with [[bug-a-help-does-not-advertise-flags-the-compiler-accepts]]: when --help omits a flag AND a bad value calls it nonexistent, both available instruments agree on a false answer.

Suggested shape

Split the two verdicts at the parse site: an option matched by name with an unacceptable value should say so and, where the set is closed and small, name the accepted values — --esp-profile already does the first half. Reserve unknown option for a name that matches no handler.

Worth a decision rather than a guess: whether the value list is printed for every closed-set option or only where it is short. --target has --list-targets already, so pointing at it beats printing the list.

Gate

make compiler/pascal26 plus the six rows above, run. Do not widen — a suite is not what checks a diagnostic's text. Note also [[chore-t-a-wikilink-to-a-ticket-that-does-not-exist-is-never-detected]]'s neighbour, face 231: unknown option: wording is asserted by grep -q in the Makefile at least once, so grep for the current wording before changing it.