[fuad-daoud]

Summary

In short: irverify decides whether a name is blank with its own copy of Unicode's list of invisible characters, and the copy was wrong for 292 code points. This makes it follow Unicode's definition: an entry of variation selectors only is now reported blank, and visible format characters such as U+0600 no longer are.

isInvisible approximated Unicode's Default_Ignorable_Code_Point as space, control, Cf and Other_Default_Ignorable_Code_Point, which is wrong in both directions.

• It missed Variation_Selector, so an entry made only of selectors was not blank.
• It counted the Cf characters Unicode takes out of the property because they render: Prepended_Concatenation_Mark, the interlinear annotation characters U+FFF9..FFFB and the Egyptian hieroglyph format controls U+13430..1343F. An entry made of those was reported blank.

It now follows the property's definition. Package unicode has a table for every term but those two ranges, which are named constants. Both callers of isBlankName change with it: the alias rule (ir/naming-alias-blank) and the string-list rules (ir/path-segment-blank, ir/set-entry-blank). docs/ir/naming.md states the new set.

The only one of these lists a compiler writes is Visibility.Only, from fixed lifecycle names, so no spec's output moves, and the IR's shape does not change.

Limitation: the two constants hold for Unicode 17.0.0, as no unicode table carries the property. TestIsInvisible_RangesWereCheckedAgainstThisUnicode fails when a Go release moves unicode.Version, so that move rechecks them against DerivedCoreProperties.txt. That file's header writes the second range as 13430..13440; U+13440 is Mn and in no added term, so subtracting it changes nothing. No open issue tracks this. #1511, closed, left a test tying a predicate to unicode.Version out of its own scope; this one pins the version string alone, and the sweep over the UCD files stays a throwaway probe, as .claude/skills/morphic-changing-the-ir/references/ir-lessons.md#sweep-a-published-property prescribes.

Closes #832

Before and after

flowchart TB
  subgraph before ["Before"]
    b0["An alias or list entry of variation selectors only"]
    b1["Passes as a name, though it names nothing"]
    b2["An entry of U+0600 or U+FFF9, which render, is reported blank"]
    b0 --> b1
    b1 --> b2
  end
  subgraph after ["After"]
    a0["An alias or list entry of variation selectors only"]
    a1["Reported blank by irverify"]
    a2["An entry of U+0600 or U+FFF9 passes as visible"]
    a0 --> a1
    a1 --> a2
  end
  before ~~~ after

Loading

What it brings: irverify's blank-name rules now hold Default_Ignorable_Code_Point as Unicode publishes it on every code point, instead of a near copy that was wrong for 292 of them. White space and controls stay invisible through their own terms, as before (#832), U+1680 OGHAM SPACE MARK among them though it renders a line.

Test plan

• CI on this head, main at 2123d296 merged in, runs make gate. On 57484050, go build ./..., go test ./internal/archtest, go test -cover ./ir/irverify at 100.0%, make vet, make lint, make lint-planted and the skills lint each pass.
• Every code point against Unicode 17.0.0's DerivedCoreProperties.txt and PropList.txt, by a throwaway in-package test: visited 1114112 ucdDiffs 0. Against the predicate on main, 260 code points are added, all Variation_Selector, and 32 removed, the characters listed above.
• Removing each term of isInvisible, one at a time, reddens a row. So does moving a bound inward, or interlinearFirst to U+FFF8. Rows for U+13440 and U+E000 keep Variation_Selector and Cf from widening to Mn, M or C, and rows for U+FFFC and U+1342F keep Cf from widening to So or Lo: each of those mutants reddens one. Narrowing the white-space term to exclude U+1680 reddens the ogham space mark row. Narrowing it to Zs reddens the line separator and paragraph separator rows. Narrowing the control term to C0 reddens delete and c1 control, and to C0 plus DEL, c1 control. Narrowing the Other_Default_Ignorable_Code_Point term with && !unicode.Is(unicode.Mn, r) reddens the combining grapheme joiner row. Three bound moves stay green and are equivalent, because the code points they add are false either way: interlinearLast to U+FFFC, egyptianFormatFirst to U+1342F and egyptianFormatLast to U+13440.
• sweep.py -b origin/main -j 2 testdata -- --verify over 236 specs: "no spec's exit code, IR or diagnostics differ". With isInvisible planted to answer true, 4 specs go from exit 0 to 2; without --verify no compile runs irverify, and the same plant shows nothing.
• TestIRVersion_PinsTheEncodedShape and TestIRVersion_PinsTheNames stay green at an unmoved IRVersion.