fix(types)!: retire ToastSchema.action from both published faces (objectui#8338) - #8540

Fixes #8338

Retires ToastSchema.action from BOTH published faces under ADR-0049 enforce-or-remove, the direction ruled at comment 5571798211. action?: never on the declaration, retirementTombstone() on the mirror.

The spine: it sits four lines above a tombstone that gives its own reason

ToastSchema.onDismiss has been an ADR-0049 tombstone since objectui#6124, and its prose reads "JSON has no function value, and the toast renderer takes ({ schema }) and never reads it." Every clause of that sentence is true of action as well. It survived only because that sweep was over TOP-LEVEL function-valued keys and this key's function is one level down — inside a required object, which is strictly worse: not "a function key is inert", but no JSON value satisfies the whole member.

The two faces shared nothing. The declaration required label and onClick; the mirror admitted SchemaNode | SchemaNode[]. One document, a green safeParse and a tsc refusal, and no spelling that satisfied both. The renderer read neither (it reads variant, title, description, duration, buttonVariant, className, buttonLabel; the file has exactly one ComponentRegistry.register call, so the zero is a reading and not a failed scan).

The parity-operator question triage ordered an answer to

Answered at comment 5572120924, and re-derived here on 3bc187b7f rather than quoted forward. The card offered two readings — the operator judged action clean, or it excused the key into the SCHEMA-NODE bucket. Neither. A third mechanism: with the mirror's input face reading unknown (the objectui#7760 defect), the wider-direction comparison could not form a reading at all, so no row existed to be right or wrong about. PR #8354 gave the recursive mirrors their declaration as a type argument, and on today's tree the key is ENTERED in KnownDrift, in WiderThanDeclared, and in WIDER_ARMS.

So the card body's section "The parity ledgers do not record it" is historical and is not quoted forward. What the retirement owed was ordinary ledger work: drain those rows and move the header's pinned figures.

The header figures, derived and not stepped by hand

Read by zod-mirror-parity.test.ts's own AST and mirror instruments — the ledger rows were removed first, the pins were run, and the numbers below are the ones those pins printed:

CONCRETE fell by 1 while arms fell by 2 — the split counts KEYS and the arm total counts ARMS, through different reducers. A hand-stepped CONCRETE was wrong by one when this was written, which is why the header now says so in as many words.

The removal is not merely allowed, it is required by the compiler: the ablation below shows tsc -p tsconfig.test.json reporting TS2322 at three sites in the ledger (Type '"feedback.zod.ts#ToastSchema"' is not assignable to type 'never') the moment the divergence goes away and the rows stay. The rows drain because the divergence is gone, and the type-level reconciliation is what says so.

Two-direction ablation, with the restore proved

One run, both legs, under a trap ... EXIT INT TERM, mutating by checking the base tree's two declarations back in and restoring with git checkout HEAD --:

ANCHOR feedback.ts   'action?: never;'  before=1 after=0
ANCHOR feedback.ts   'action?: {'       after=1
ANCHOR feedback.zod  tombstone          before=1 after=0
ANCHOR feedback.zod  z.union arm        after=1
MUTATION REACHED DISK - both tombstones removed, both pre-retirement declarations present
ABLATED_VITEST_EXIT=1
ABLATED_TSC_EXIT=2
RESTORE PROVED: both blobs byte-identical to HEAD, `git diff HEAD` empty
RESTORED_VITEST_EXIT=0
RESTORED_TSC_EXIT=0

Ablated, the runtime pins go red on exactly the shapes that used to parse:

AssertionError: still accepts {"type":"button","label":"Undo"}: expected true to be false
AssertionError: still accepts [{"type":"button"}]: expected true to be false
AssertionError: expected 'invalid_union' to be 'invalid_type'
AssertionError: expected 'Invalid input' to be 'Action button'

Ablated, tsc -p tsconfig.test.json reports:

toast-button-keys.test.ts(307,5): error TS2578: Unused '@ts-expect-error' directive.
toast-button-keys.test.ts(331,10): error TS2344: Type 'false' does not satisfy the constraint 'true'.
zod-mirror-parity.test.ts(2370,14): error TS2322: ... not assignable to type 'never'.
zod-mirror-parity.test.ts(2459,14): error TS2322: ... not assignable to type 'never'.
zod-mirror-parity.test.ts(2478,14): error TS2322: ... not assignable to type 'never'.

Only ONE of the two @ts-expect-error directives goes unused under ablation, and that is correct: the old declaration accepted the { label, onClick } object but still refused a node, so the second directive keeps firing. Stated because a report that claimed both would be claiming more than was measured.

The mirror also admitted BARE PRIMITIVES — re-derived, and larger than first read

SchemaNodeSchema is z.union([BaseSchemaCore, z.string(), z.number(), z.boolean(), z.null(), z.undefined()]) (zod/base.zod.ts:84), so the old z.union([SchemaNodeSchema, z.array(SchemaNodeSchema)]) accepted far more than two node shapes. Re-derived by re-forming the old spelling around the SHIPPED SchemaNodeSchema rather than a hand-rebuilt one, with a non-vacuity control (the two envelopes must disagree on the node case, else the probe compares a schema with itself — measured old=true new=false):

The primitive row is the LARGER half of this narrowing and the easy one to miss; the changeset's table now carries it. RETIRED_MIRROR_SHAPES gained 'Undo', 1 and null, and the list's MIXEDness is now pinned in its own right: vitest 4.1.10 spreads an it.each case only when cases.every(Array.isArray), so the list arm arriving as ONE argument had been resting on an undeclared property of the data. The three added cases show up as three added tests (2795 -> 2799 with the mixedness pin), which is the reading that they were passed whole rather than spread.

The tombstone message was respelled, and the emitted string is asserted unchanged

scripts/check-widening-tells.mjs's T2 arm matches BARE_STRING_ELEMENT — an indented bare quoted string that ENDS a line — so the tombstone's first message fragment, sitting alone on its own line, read as a closed set gaining a member and this NARROWING reported one widening tell. Exactly one, because every continuation carries a leading + , which is itself the proof of the diagnosis. The fragment moved onto the retirementTombstone( call line; no message TEXT changed.

Asserted, not eyeballed: the emitted .describe() string is read off a rebuilt dist before and after and hashed — len 732, sha256 10b0b510c8b4ece4ab594373f870644dcefb9816b9bc92608b0791a060190fcc, identical, with the probe exiting non-zero on any drift; the parse-time issue message still equals .describe() and the code is still invalid_type. Self-scan: the same BARE_STRING_ELEMENT regex over every ADDED line of the whole diff now finds 0 hits, with a synthetic 'destructive', as the control that the regex fires.

retirementTombstone(), and why NOT the neighbour one line down

Triage's model was "照 onDismiss 一模一样的处理,四行之外就是范本". Measured, the neighbour is handlerKeyRefusal('onDismiss', 'retired', 'Dismiss handler') — the handler-key variant, which reports custom and whose remedy sentence is fixed at "author behaviour as a NODE TYPE". action is not a handler key: it is a value key whose nested member was a function. It takes retirementTombstone() (invalid_type, and a guidance string this site writes for itself), which is what triage actually named. Copying line 77 verbatim is the obvious move and it is the wrong one. tombstone.zod.ts is called, not edited.

What replaces it: nothing, and it was never fulfilled

Not "not yet supported" and not a pointer at a future shape. objectui#6250's close moved all seven toast demos off an in-toast action entirely; an in-toast action button is a capability expansion with zero runtime today and would need its own card. The tombstone points at the keys that DO run: raise the toast from the node itself and label its trigger with buttonLabel / buttonVariant.

Repo census: of the 7 authored toast nodes in tracked JSON, 0 carry action (control: the same walk finds all 7), and no source file authors one ON A toast NODE. Stated that narrowly on purpose — app-shell/src/chrome/notificationToast.tsx:76 and chrome/toast-helpers.ts:71 DO write action: { label, onClick }, for sonner's runtime API, a different interface this key never fed. The changeset carries the same correction: a measured zero must not be quoted wider than what was counted.

Not EmptySchema.action

The same word lives 69 lines below in the same file (feedback.zod.ts:152), on a sibling schema, with the deliberately opposite disposition (objectui#7105, PR #8330) and its own comment block explaining that it refuses the { label, onClick } shape this key spelled. It is untouched. Two different interfaces, on purpose.

Verification

Every exit code captured by redirect BEFORE any pipe; every verdict quoted from the gate's own line, never from a bare $?.

vitest alone would be a FALSE GREEN for this ledger: its type-level reconciliation is judged by tsc -p tsconfig.test.json, and --listFiles confirms that project compiles both zod-mirror-parity.test.ts and toast-button-keys.test.ts — so the green is a measurement and not an absence.

All figures re-derived on e8d984461, the final commit; the ratchet family (packages/types tests + all three tsc projects) was re-run on that exact head and is exit 0.

验收备注

• Filed: objectui#8539 — content/docs/components/feedback/toast.mdx:40-44 still teaches action?: { label, onClick } in its plaintext schema fence. Outside this PR's declared file surface, and measured to be read by NO gate (check-doc-component-types.mjs states in its own header that it judges the type string only). The snippet was already uncopyable before this PR — no JSON document satisfied it — so this is a pre-existing teaching defect whose failure mode this PR changes. Same shape as objectui#7340 and objectui#7239, both of which were filed as their own docs cards rather than ridden into the tombstone PR.
• noted, not filed: packages/components/src/__tests__/empty-action-declared-7105.test.tsx:149 says in a comment that { label, onClick } "is ToastSchema.action's shape" — present tense, now historical. Cosmetic prose inside another card's pin file (objectui#7105 / PR feat(types,components): declare EmptySchema.action as SchemaNode and enforce it #8330), which this PR is ruled not to touch. Whoever next edits that pin will meet it.
• noted, not filed: the toast page's fence also omits position, buttonLabel, buttonVariant and the onDismiss tombstone. A hand-maintained subset; recorded on objectui#8539 as an observation and deliberately not folded into its ask. Successor: whoever takes objectui#8539.
• Draft, and it stays draft: SKILL.md:643 puts *.zod.ts on the contract path limb, so the pre-enqueue in-seat contract review is the PM's to run. This seat does not enqueue, does not flip to ready, and hangs no label the dispatch did not name.
• Bundle Analysis is red here and is ⛔ NOT this PR's: main itself has been red on that check since f76f4362, four consecutive merges, with the framework chunk over its per-chunk ceiling. Tracked at objectui#8542. Nothing was trimmed for bytes and PER_CHUNK_GZIP_CEILINGS is untouched.
• Clause-②: no — the retirement NARROWS both faces. Re-checked against the landing: the mirror's accept set loses the node union and gains nothing, and the declaration goes from an object type to never. Nothing widens.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CZY49skxUBYyJcdnTcYPrE

Generated by Claude Code