Source-scoped overlay invalidation proof must separate recompute selectivity from subscriber fan-out

Problem

The perf lane kept treating source-scoped invalidation like an abstract future promise.

That was too fuzzy. We already had a live React locality benchmark, but it did not explicitly prove whether source dirtiness stayed selective, and it did not distinguish that from the separate question of whether touched store subscribers still fan out too broadly.

Symptoms

• source-scoped overlay invalidation stayed marked as “no command-backed live proof yet” in the control docs.
• The existing rerender lane proved selection breadth, deep ancestor locality, decoration toggle breadth, hidden Activity, and annotation-backed widget churn, but not the exact dirtiness-class split.
• It was too easy to hand-wave with “overlay locality is green” or “source invalidation is still missing” depending on the mood of the doc writer.

What Didn't Work

• Treating decorationSourceToggleBreadth as if it answered the whole source-invalidation question. It only proved one external refresh path.
• Treating the absence of a dedicated row as proof that no owner existed.
• Treating rerender breadth and recompute breadth as the same thing. They are not.

Solution

Extend the live rerender-breadth.tsx owner with one explicit sourceScopedInvalidation row instead of inventing a second benchmark family.

The new row creates three projection stores on the same editor:

• a selection-dirty store
• a text-dirty store
• an external-dirty store

Then it drives three operations and records recompute deltas separately:

const selectionStore = createSlateProjectionStore(editor, deriveSelectionRanges, {
  dirtiness: 'selection',
  sourceId: 'selection-source',
})

const textStore = createSlateProjectionStore(editor, deriveTextTailRanges, {
  dirtiness: 'text',
  sourceId: 'text-source',
})

const externalStore = createSlateProjectionStore(
  editor,
  () => deriveExternalRanges(externalActiveRef.current),
  {
    dirtiness: 'external',
    sourceId: 'external-source',
  }
)

Fresh kept result from bun run bench:react:rerender-breadth:local:

• selection change:
  • selection recompute count: 1
  • text recompute count: 0
  • external recompute count: 0
• text edit:
  • selection recompute count: 0
  • text recompute count: 1
  • external recompute count: 0
• external refresh:
  • selection recompute count: 0
  • text recompute count: 0
  • external recompute count: 1

That proves the dirtiness classes are selective.

The same row also exposes the still-open part:

• selection store subscriber rerenders: left 1, right 1
• text store subscriber rerenders: left 1, right 1
• external store subscriber rerenders: left 1, right 1

So the current store layer does decide which store needs recompute, but once a store snapshot changes, both runtime-id subscribers on that store still rerender together.

Why This Works

It splits one muddy perf argument into two crisp questions:

1. Does the runtime recompute only the touched source class?
2. Once that store changes, does subscriber delivery stay local inside the store?

Before this row, those questions were getting mashed together.

Now the answer is explicit:

• recompute selectivity: yes
• per-store subscriber locality: not yet

That is a much better checkpoint than either fake green or fake missing.

Prevention

• When adding overlay perf proof, separate recompute metrics from rerender metrics. One without the other invites bullshit conclusions.
• Prefer extending the existing owner command when the new question is adjacent to the same runtime family.
• Do not mark a perf lane “missing” once a command exists just because the result is not fully green.
• Do not mark a perf lane “closed” just because recompute counts are selective; subscriber delivery may still be too broad.

Related Issues

• Slate React rerender breadth is mostly local and useSlate is the only broad hook left
• Overlay perf coverage must include annotation-backed widget churn
• Slate DOM + React Tranche 5/6 Execution