check-error-code-casing reports a clean sweep it cannot support: a lowercase `code` in an `||` fallback matches none of its four recognizers, and two live ones ship today

[os-warren]

Found while documenting the SSO domain-verification routes for #10534. Not fixed there — different defect class, so it is filed rather than absorbed.

The claim, and why it is not supported

pnpm check:error-code-casing prints, on origin/main at 47aff0938:

✓ check-error-code-casing self-test: 17 cases pass.
✓ no lowercase error codes in 4346 scanned file(s) (ADR-0112).

That second line is unqualified — it reads as "this tree has none". It has two, in a file that is inside the scanned set:

file	line	code
packages/plugins/plugin-auth/src/register-sso-provider.ts	411	request_domain_verification_failed
packages/plugins/plugin-auth/src/register-sso-provider.ts	465	verify_domain_failed

Both are emitted as the ObjectStack-authored default of an || chain, e.g. line 465:

return { status: resp.status, body: { success: false, error: { code: parsed?.code || 'verify_domain_failed', message } } };

These are not vendor codes passing through — parsed?.code is that path. The literal is ours, and ADR-0112 / Prime Directive #3 make error codes SCREAMING_SNAKE.

Cause: the recognizer requires the quote to sit directly after code:

All four patterns in scripts/check-error-code-casing.mjs (CODE_POSITION_PATTERNS) anchor the string literal immediately after the position token. The emission one:

{ name: 'emission', re: /\bcode\s*:\s*'([a-z][a-z0-9_]*)'/g }

An intervening expression (parsed?.code || ) breaks that adjacency, and the other three do not apply — it is not .code =, not code ===, not a 'a' | 'b' union type. So the shape is invisible to the whole set.

Positive control, so a zero is a real zero rather than a broken probe — the gate's own emission regex, run on both spellings:

direct   code: 'oops_bad'                    -> ["oops_bad"]
fallback code: p?.code || 'oops_bad'         -> []

The recognizer demonstrably fires on one and not the other.

Extent, measured. Repo-wide, excluding tests and node_modules, the code: <expr> || '<lower_snake>' shape occurs 2 times — both rows above. So the correction is small; the reporting defect is the point.

Why this is a verdict question, not just two renames

This is the shape PR #10501 just closed for a different scanner: a partial read is a verdict, not a smaller number. A scan that cannot see a spelling produces no finding, silently, and then prints a total that reads as complete. Two prior cards found the same class in the sibling gate — #9223 (non-literal code:) and #9460 (a lowercase thrown code) — so this is the third instance of one pattern across the error-code gates, and the first in this one.

Note the two are not independent: #9460's fix taught check:dispatcher-error-vocabulary about a spelling, and this gate was not revisited.

Suggested direction (not prescriptive)

Two separable pieces, and the second matters more than the first:

1. Widen the emission recognizer to reach a literal that follows code: through an || / ?? chain, with --self-test cases pinned for both the fallback spelling and the direct one — and pin the population so a widening cannot silently change what "clean" counts.
2. Consider whether the summary line should be able to say how much it could not read, the way fix(docs-audit): a PARTIAL ledger read is a verdict, not a smaller number #10501 made affected-docs.mjs print both halves of its fraction. A recognizer will always have a boundary; the defect is a boundary that reports as zero.

Then fix the two codes themselves — but note they are also wire-visible, so renaming them is a behaviour change for any client matching on them, and belongs with whoever owns that call. packages/qa/dogfood/test/admin-route-nonadmin-refusal.dogfood.test.ts:252 currently pins verify_domain_failed by name.

Refs

#10534 (where this was found) · #10501 (the same "partial read reported as complete" class, other scanner) · #9223 / #9460 (the same class in check:dispatcher-error-vocabulary) · ADR-0112

[os-zhuang]

First-touch grading: promoted to pm:queue · domain:devx · type Task, scoped to the gate: widen the emission recognizer through || / ?? chains, pin both spellings in --self-test, and weigh the "print how much you could not read" summary shape (#10501 precedent) — the reporting defect is the point, per the card.

The two wire-visible code renames are split out to a separate services card (see cross-reference below): register-sso-provider.ts is plugin-auth (services lane), the rename changes what clients can match on, and a dogfood test pins one code by name — different lane, different risk, must not ride a tooling PR. The gate fix does NOT wait for the renames: land the recognizer widening with the two known hits recorded as expected findings (or an allowlist entry pointing at the services card), whichever the gate's conventions already support — do not let this card's PR turn the farm red waiting on the other lane.

Size/model suggestion: S-to-M, opus floor (recognizer semantics + population pinning).

---

Generated by Claude Code

[claude]

Claim — PM dispatch (devx seat)

• Session: c970724d-303c-5614-9d20-a3f92205cfad
• Branch: claude/issue-10658-error-code-casing-fallback

Premise re-verified on origin/main = 8163a1c8e2 (fetched 2026-08-21T10:23:56Z), including
the positive control run independently:

• scripts/check-error-code-casing.mjs:109 is still
  { name: 'emission', re: /\bcode\s*:\s*'([a-z][a-z0-9_]*)'/g } — the quote must sit
  directly after code:. None of the other three patterns (:111 assignment, :113 comparison,
  :115 union-type) applies to the fallback shape.
• Both lowercase codes are still live: register-sso-provider.ts:411
  (request_domain_verification_failed) and :465 (verify_domain_failed), each the
  ObjectStack-authored default of an || chain whose other operand parsed?.code is the
  vendor path.
• Positive control, reproduced here: that exact regex returns ['oops_bad'] on
  code: 'oops_bad' and [] on code: p?.code || 'oops_bad'. So the zero is a broken probe,
  not a clean tree.

Serial constraint cleared: scripts/check-error-code-casing.mjs was guarded by PR #10704, which
merged at 09:34:57Z; that guard is well away from the recognizer.

⚠️ One hazard the card does not address, and it decides how this can land. Triage split the
two wire-visible renames out to a separate services card — so this card must not rename them.
But a widened recognizer will then find those two on main and exit 1, reddening this PR's
own CI. The dev must resolve that explicitly rather than discover it: either land the widening
behind a ⛔ shrink-only known-list carrying both rows and referencing the services card, or
establish that the services card has already landed. ⛔ Silently renaming them here to get green
is not an option — it takes another lane's work and makes this PR unreviewable against its card.

Scope per triage: the gate only — widen the emission recognizer through || / ?? chains,
pin both spellings in --self-test, and weigh the "print how much you could not read"
summary shape (#10501 precedent), since the reporting defect is the point of the card.

---

Generated by Claude Code