[finding] the dispatcher gate's resolver anchors on `const`, so a `let`-held error code is reduced by nobody even when every branch is a literal

[claude]

Found while closing #10897 (teaching check-error-code-casing to read the
our-default slot in a code local's initializer). Filed rather than absorbed:
it is the OTHER gate's resolver, it is a behaviour change to a deliberate bound,
and #10897 was explicitly scoped to the recognizer.

The shape

declaredInitializers in scripts/check-dispatcher-error-vocabulary.mjs
anchors its head regex on const alone:

const head = new RegExp(
 `\\b${exported ? 'export\\s+' : '(?:export\\s+)?'}const\\s+${name}\\s*(?::[^=;\\n]+)?=\\s*`,
 'g',
);

So a code held in a let reduces to nothing. resolveConstant answers null,
the codehelper branch reads that as "unreducible" and takes the documented
runtime-value bound — no site and no unresolved — even when the value is a
plain string literal sitting right there in the source.

That is not the bound #9568 wrote. Its words: "a value a source scan cannot
evaluate is out of reach; a value spelled out in every branch is not, whichever
shape it arrives through." A let initialized to a literal is spelled out. It
is out of reach only because of how the anchor is spelled.

Measured, with a positive control on each side

Driving deriveSites and findViolations on origin/main at 58563be069,
one fixture per row, each a complete export function that stamps err.code = code and throws:

shape	check:dispatcher-error-vocabulary	check:error-code-casing
let code = 'REASSIGN_FALLBACK'; if (x) code = 'REASSIGN_OTHER';	no site, no unresolved	0
let code = 'REASSIGN_FALLBACK'; if (x) code = p?.code || 'reassigned_lower_failed';	no site, no unresolved	0
control: const code = 'local_direct_failed'	1 site (assignconst)	0
control: const code = flag ? 'tern_lower_a' : 'tern_lower_b'	2 sites (assignconst)	0
control: const code = p?.code || 'reassigned_lower_failed'	no site	1 (local-fallback, after #10897)

Rows 3 and 4 are the control that makes rows 1 and 2 readable rather than a
broken probe: the identical shape spelled const does reduce, lowercase
included, and does produce sites. Row 5 is the casing-gate control on the other
side. Row 1 is the sharpest one — both of its values are literals, both
unregistered, and no gate says a word.

Two separable halves

1. let / var declarations. The cheap half, and the one that matches
   [finding] check:dispatcher-error-vocabulary still cannot see a code held in a local ternary — sys-metadata-repository's live 403 NOT_CREATABLE / NOT_OVERRIDABLE #9568's stated intent: widen the declaredInitializers anchor. The reason to
   be careful rather than quick is that a let can be reassigned, so the set
   of values a let holds is not simply its initializer — reducing it to the
   initializer alone would report a value the program may not stamp, which is
   the same wrongness the ALL-OR-NOTHING rule ([finding] check:dispatcher-error-vocabulary still cannot see a code held in a local ternary — sys-metadata-repository's live 403 NOT_CREATABLE / NOT_OVERRIDABLE #9568) exists to refuse. The
   honest widening probably reduces a let only when the file contains no
   reassignment of it, and reports it as unresolved when it does — never
   silently, which is the bound this gate holds itself to everywhere else.

2. Bare reassignment (code = p?.code || 'lit', no declarator) — row 2.
   Harder: no declaration to anchor on, and a textual scan has no scopes. Worth
   recording, not necessarily worth closing. Note [finding] check-error-code-casing's ||/?? fallback recognizer anchors on the STAMP SITE only, so our authored default in a const code = … initializer is still seen by no gate #10897's new local-fallback
   recognizer deliberately requires a declarator, so the casing gate does not
   reach this one either.

Extent today: not swept

I did not sweep the tree for live victims — that measurement belongs with
whoever takes this, since the right population depends on which half is being
closed. What is known: 47 code-named local declarations exist under
packages/** (measured for #10897's blast radius), and at least one of them is
a let (packages/metadata-protocol/src/protocol.ts:6551,
let code: unknown | null =). Whether any holds an unregistered literal is the
open question.

Search note

The issue-search endpoint is unavailable from an agent seat here ("sessions are
bound to their configured repositories"), so the duplicate check behind this
card is a title grep over the 100 most recently created open issues, not a full
search. Worth a second look before anyone starts.

Refs

#10897 / PR #10914 (where this was found) - #9568 (the ALL-OR-NOTHING reduction
and the bound this contradicts) - #9460 (the local-variable bound it refines) -
ADR-0112 D1

Generated by Claude Code

---

Generated by Claude Code

[huangyiirene]

Triage (first-touch): promoted finding → pm:queue + domain:devx; rationale: measured with positive controls on both sides, and half 1 restores #9568's own stated bound ("a value spelled out in every branch is not out of reach") rather than widening it. Scope: half 1 only — widen declaredInitializers to let/var, reducing a let only when the file contains no reassignment of it, and reporting unresolved (never silent) when it does. Half 2 (bare reassignment, no declarator) stays recorded here, not queued — reopen it as its own card if a live victim is found. The extent sweep (47 code-named locals, which hold unregistered literals) belongs to the executing dev and its result goes in the PR body. Dedupe re-checked at triage: #10897 (casing-gate recognizer, in-flight) is the sibling, not a duplicate — different gate, and this card's rows 1–2 are invisible to both gates today. Type Task. Serial note: same script family as in-flight #10897/PR #10914 — serial after it lands. Size/model suggestion: M, mode:subagent, model: opus.

---

Generated by Claude Code