Chore(CI): repair spell checking and fix the typos it finds

[matthewmcneely]

Makes spell checking able to run, and fixes the 36 real typos it turns up.

This PR alone does not switch spell checking back on. GitHub Actions is disabled repository-wide (actions/permissions reports "enabled": false), and ci-spell-checking is additionally marked disabled_manually. Both are repository settings, not files, so they need an admin to flip them. See "Before this takes effect" below, which includes a hazard worth reading first.

The problem

Three separate things were wrong, and each one alone was enough to make the check vacuous.

The path filters pointed at a directory that no longer exists. Both workflows filtered on ./content/, which has not existed since the Docusaurus migration. any_changed was therefore always false and every step was skipped. The runs still reported success, which is why this went unnoticed. The last real run of ci-spell-checking was 2025-11-05.

Actions was later disabled entirely. The last run of any of these workflows was 2025-12-04. Nothing has run since.

Repointing the paths is still not enough, because the check then fails immediately on existing content. Two further causes:

Vale.Terms was reading the vocabulary as a casing authority. .github/styles/config/vocabularies/Dgraph/accept.txt is a spelling accept-list, but Vale.Terms enforces the exact casing of every entry it finds there. Because the file lists both GraphQL and graphql, and lists Cluster without cluster, the result was 1,378 error-level false positives across the corpus:

Message	Count
Use 'graphql' instead of 'GraphQL'.	652
Use 'Cluster' instead of 'cluster'.	287
Use 'Dgraph' instead of 'dgraph'.	118
...and 12 more patterns	321

write-good was enabled. Its E-Prime rule flags every use of "is", "are", and "be", which is unworkable for reference documentation: 25 alerts in a 3-file sample. It is a prose-style linter covering ground documentation-style-guide.md already covers, by a different standard.

What changed

• .github/workflows/ci-spell-checking.yml — paths repointed at the four next-content directories, *.md and *.mdx. Also pins the Vale version so a new release cannot silently change what CI accepts, moves to actions/checkout@v4 with the history changed-files needs, and passes the file list through the environment rather than interpolating it into the shell.
• .github/workflows/deployments.yml — deleted. It ran reviewdog misspell, which Vale supersedes, and its filename described a deployment it never performed. CloudFlare Pages handles deployment through its GitHub integration, not a workflow.
• .github/.vale.ini — Vale.Terms off with a comment explaining why, write-good dropped, mdx added to the file glob.
• accept.txt — 145 entries added and the file sorted, so future diffs stay readable. The additions are identifiers (backupId, forceFull, taskId), technical terms (backpressure, linearizability, snapshotting, varints), product names, abbreviations (PVCs, SSTs, vCPUs), and a few deliberate oddities in sample data: goe and sourc are stemmer-output fragments in a tokenizer example, and Un nouvel is French sample data in an RDF example.

The typos it found

20 misspellings, including Aplha → Alpha in the glossary, transactoins → transactions, Retreive → Retrieve, prediate → predicate, forfull → fulfill, wouild → would, parctices → practices, attriubutes → attributes, and Minmizes → Minimizes in two link labels.

16 duplicated words, including will lead the the declaration → to the, must be equal the the JWT claim → to the, and what your app needs to to → to do.

All are fixed in docs/, docs-graphql/, and docs-learn/, and backported to version-v25.4 in both versioned trees, since those are live on the site today. Older snapshots are left as the historical record they are.

Scope

Linting covers docs/, docs-graphql/, docs-ratel/, and docs-learn/ only. The *_versioned_docs/ snapshots are frozen copies that are not edited in place, and a version cut copies the next content forward already checked, so linting them would double every alert for no benefit. It also means a release PR like #763 will not drag 227 snapshot files through the linter.

Before this takes effect

Re-enable Actions for the repository (Settings → Actions → General), then re-enable the ci-spell-checking workflow, which is separately marked disabled_manually. Both are repository settings rather than files, so neither can be changed by this PR.

Stale bot, handled here

stale.yml runs nightly at 01:30 UTC and had no exempt labels configured, so re-enabling Actions after this long a quiet period would have marked a backlog of still-wanted work as stale. This PR adds exempt lists:

• Issues: bug, accepted, RFC, help wanted, good first issue, investigate, triage:delayed, reviewing
• PRs: accepted, RFC, reviewing, triage:delayed, cherry-pick

Dependabot PRs are deliberately left unexempt. The next bump supersedes them, so closing is the right outcome.

Measured against the open backlog, every item that would currently qualify as stale is now handled:

Category	Count	Outcome
Issues labelled bug (#700, #707, #713, #730)	4	Exempt
Contribution PRs (#755, #728, #708, #699)	4	Labelled reviewing, now exempt
Dependabot PRs	12	Marked stale, as intended

The four contribution PRs carried no labels, so no exempt list could reach them. They have been labelled reviewing separately from this PR.

Testing

Since CI cannot run yet, this was verified locally with Vale 3.7.1, the version the workflow now pins:

• All four content directories: 0 errors, 0 warnings, and 0 suggestions in 293 files, exit code 0.
• The CI invocation simulated against this PR's own 30 changed files: also clean, exit code 0.
• npm run build passes, confirming the two corrected link labels still resolve.

The workflow YAML parses, but it has not executed. That is the one thing here I could not verify end to end.