csstype `Property.*`/`Globals` value-unions enumerated instead of mapped to `string` when reached through `DeepPartial<>` — 872 giant `@unboxed` bodies + CSS-keyword constructors leak into shared namespace (blend 0.0.38-beta.1)

[jagguji]

Summary

csstype Property.* / Globals value-unions (Color, BackgroundColor, the
-moz-initial | inherit | initial | revert | unset … globals, the full 148 CSS
color keywords, etc.) are enumerated into giant @unboxed variants instead
of collapsing to the precise string mapping the extractor already knows how to
apply (isCssType, extract.mjs:1092). This fires when the csstype value is
reached through blend 0.0.38-beta.1's new ComponentTokenOverrides = DeepPartial<ComponentTokenType> wrapper — the same trigger as #205, but a
different symptom on a different code path.

Bindgen held constant, only the blend version bumped 0.0.38-beta.0 → 0.0.38-beta.1:

• 872 emitted identifiers ≥ 200 chars in ContextSharedTypes.res alone (beta.0: 0); longest emitted line ≈ 12,071 chars.
• ~3,076 CSS-keyword constructor lines (@as("aliceblue") Aliceblue, @as("-moz-initial") MozInitial, @as("currentcolor") Currentcolor …) added.
• The collision report's "left as-is" ambiguous constructors jump 162 → 397 — CSS keywords (Aliceblue, AccentColor, Beige, Blueviolet … +241 more) now leak into the shared/collision namespace of ContextSharedTypes.
• The disambiguation table degrades from readable names (CenterConfigAlign, LeftConfigSide) to monstrosities like CenterMozInitialOrInherit…OrCurrentColorOrContextComponentTokensTAGSSmMinWidthConfig.

Not the same as the already-filed issues

• Anonymous large @unboxed union names bloat to 2000+ chars (CSS Property.* mixes in blend 0.0.38) #200 / Fix #200: cap bloated anonymous @unboxed union names (prefix + shape hash) #201 (closed): capped the name length of these anonymous unions to prefix + hash. That fix is cosmetic by its own commit message ("the emitted @unboxed type body is unchanged"). The giant enumerated body and the CSS-keyword constructor leakage into the collision namespace both remain. This issue is about eliminating them at the source, not shortening their names.
• Record token fields silently widen to string when a DeepPartial wrapper deepens the graph (blend 0.0.38-beta.1: +310 record→string regressions, source unchanged) #205 (open): record token fields → string via MAX_DEPTH + healGhostRecords rejection. Same upstream trigger (DeepPartial<ComponentTokenType>), but a structural-record path. This issue is the csstype-value-union path: fields that stay typed but expand every CSS keyword instead of being string.

TS input → current output → expected output

TypeScript (via DeepPartial<ComponentTokenType>):

// a token leaf whose type is a csstype property union
minWidth?: Property.MinWidth<string | number>   // = "-moz-initial" | "inherit" | … | (string & {}) | number | …

Current ReScript output (beta.1):

@unboxed type centerMozInitialOrInherit…OrCurrentColorOrContextComponentTokensTAGSSmMinWidthConfig =
  | @as("-moz-initial") MozInitial
  | @as("inherit") Inherit
  | @as("aliceblue") Aliceblue
  | …  /* ~150 CSS keyword constructors */
  | Currentcolor
  | ContextComponentTokensTAGSSmMinWidthConfig

Expected (what isCssType already produces for a directly-referenced csstype value):

minWidth: string,   // csstype Property.* → string (precise, not a loose fallback)

Suspected root cause

isCssType (extract.mjs:1092-1099) recognises csstype only by the
declaration's source-file path (/csstype/):

const sym = type.aliasSymbol || (type.getSymbol && type.getSymbol()) || type.symbol
const decl = sym && sym.getDeclarations && sym.getDeclarations()[0]
const file = decl && decl.getSourceFile().fileName
return !!(file && /[\\/]csstype[\\/]/.test(file))

When the value is reached through the homomorphic mapped type
DeepPartial<ComponentTokenType>, the checker hands back a synthesized /
apparent union of string-literal + globals types whose aliasSymbol /
declaration provenance no longer points at the csstype package file — so
isCssType returns false, the csstype→string shortcut is skipped, and the
union falls through to full member enumeration (→ #200's name bloat, cosmetically
capped by #201, plus the constructor leakage this issue is about).

Directly-referenced csstype values (not wrapped) still collapse to string
correctly — which is why beta.0 had 0 of these and beta.1 has 872.

Fix direction

Make csstype detection structural, not purely path-based, so it survives
mapped-type wrapping — e.g. recognise a union that is dominated by the csstype
Globals set (-moz-initial | inherit | initial | revert | revert-layer | unset)
and/or the CSS named-color vocabulary, optionally with a (string & {}) open arm,
and map it to string. Alternatively, unwrap DeepPartial<> / homomorphic
mapped types back to the source property type before classifying, so the original
aliasSymbol (and its csstype declaration file) is still visible to isCssType.

Collapsing at the source removes the 872 giant bodies entirely (not just their
names), stops the CSS-keyword constructor leakage into ContextSharedTypes, and
should also shrink the type graph enough to relieve part of #205.

Repro

npx @juspay/rescript-bindgen --pkg @juspay/blend-design-system@0.0.38-beta.0 --out /tmp/b0 --webapi --report
npx @juspay/rescript-bindgen --pkg @juspay/blend-design-system@0.0.38-beta.1 --out /tmp/b1 --webapi --report
# ≥200-char identifiers: 0 in b0, ~872 in b1 (ContextSharedTypes.res)
grep -oE '[A-Za-z0-9_]{200,}' /tmp/b1/ContextSharedTypes.res | wc -l
grep -oE '[A-Za-z0-9_]{200,}' /tmp/b0/*.res | wc -l

blend 0.0.38-beta.1 is already pinned in the benchmark, so this reproduces
in-repo. First observed in the downstream regen
blend-rescript#147, which
was generated against the last published bindgen (1.4.0-beta.1, pre-#201).

[jagguji]

Resolved by #209 (Fix #205/#206: recover csstype values + token records reached through DeepPartial).

Verified on the current post-#209 blend 0.0.38-beta.1 baseline — the csstype value-union enumeration this issue describes is gone:

Symptom (pre-#209)	Post-#209 baseline
~3,076 CSS-keyword constructor lines (@as("aliceblue") Aliceblue, @as("-moz-initial") MozInitial, @as("currentcolor") Currentcolor …)	0
CSS globals set enumerated (-moz-initial / unset / revert-layer / currentcolor / transparent)	0
148 CSS color-keyword constructors leaking into ContextSharedTypes collision namespace	0
872 giant @unboxed bodies ≥200 chars in ContextSharedTypes.res alone; longest line ≈12,071 chars	218 long lines total across ALL files, and they are HTML tag-name polyvar unions (fromTag: [#"symbol" | #"object" | …]) + legitimate blend enums ("barnes-hut", "archimedean", "all-scroll") — not csstype

#209 added the structural isCssType fallback (the -moz-initial Globals fingerprint, which recognises a csstype value however it was reached, incl. through DeepPartial<>) plus isForeignArm/unionStringLiterals so a union that also carries a genuine foreign arm is never wrongly collapsed. That routes the DeepPartial-reached csstype value through the isCssType → string mapping — exactly the fix this issue asked for.

The only CSS-global residual is 2 @as("inherit") constructors, and both are legitimate small enums (Monaco's wordWrapOverride1 = "off" | "inherit", a Highcharts translation enum), not the csstype globals set.

The genuinely-remaining follow-ups from #209's review are tracked separately: #208 (deep-record heal threshold — partially-degraded token ghosts) and #210 (the (string & {}) open-arm → { ...JsxDOM.domProps } bag, cosmetic).