Use American English throughout

[mattpodwysocki]

Spellings were mixed across 28 files — mostly my own writing. honoured, behaviour, recognise, serialise, neighbour, licence, centred, whilst, sanitised, defence and friends.

Nearly all of it is comments and doc prose, but some is text a user actually sees, so this isn't purely cosmetic:

Where	Was
auth status	unrecognised prefix
a bad token	Mapbox does not recognise the token it would be given.
completion, tilesets-cli (×2)	Warning: `--output` is not honoured by …
a result that won't serialize	{"code":"error","message":"unserialisable error"}

Second commit: the convention is now a test

There's no AGENTS.md in this repo, so the instruction went into CONTRIBUTING.md — and specifically into the section that already lists four rules "the compiler holds rather than a reviewer". A spelling nobody checks is exactly what drifts across 27 files before anyone notices, so prose_is_american_english in tests/source_guards.rs holds it instead.

It earned its place immediately. My sweep used a word-boundary regex; the guard found 16 lines it had missed, every one real:

• eprint_labelled, labelled_lines and five test names around them — identifiers, where \b never matches inside labelled_lines
• mislabelled, Serialising — the British word as a prefix or suffix of a longer one
• "unserialisable error", which the CLI prints
• Honouring at the start of a sentence

That's the argument for the test over the instruction, made by the test on its first run.

Two deliberate exemptions. Fenced code blocks, because sample output and captured API responses are quoted rather than written — docs/commands.md is full of both, and a British place name in a geocoding result is not a misspelling; rewriting it would make the doc misquote the API. And cancelled, for the reason below.

Also: the table of spellings to avoid is itself a list of spellings to avoid, so the check skips a line that is nothing but two quoted strings and a comma — narrow enough to excuse a data row and not a sentence.

Verified both ways, not just by a green run:

# British spelling in prose
src/output.rs:88: recognis -> recogniz          → FAILED

# the same word inside a fenced block
test result: ok. 1 passed

cancelled is deliberately left alone

CliError::new("cancelled", "Cancelled — nothing was sent.")

That string is the machine-readable code in the JSON error contract, documented in docs/commands.md's status table and asserted in tests/non_interactive.rs, so anyone scripting against us may be matching on it. Renaming it is a breaking change, not a spelling fix. My recommendation is to do it, before 1.0 while it's cheap, but as its own PR with its own changelog entry — happy to open that.

Scope

• openapi/ and custom-openapi/ excluded, and the guard excludes them too: their descriptions reach users as help text, but the words are the API teams', mirrored from upstream, and an edit wouldn't survive a regenerate. Both are clean today.
• No changelog entry. Error message text isn't one of the promises CONTRIBUTING.md lists — command names, flags, the two output modes and the exit codes are. (This also keeps the PR off the same ## Unreleased lines as Document installing without the install script #26.)

597 tests, cargo fmt --check and cargo clippy --locked --all-targets clean.

[zmofei]

@mattpodwysocki I’ll merge this one and rebase the other PRs onto main, since I noticed some duplicate changes across the PRs.