Review follow-up: toolbar width near-twins

Scope: the $toolbar_button_raw_width / $toolbar_button_width pair named in fable-review.md P5-3. Both duplicated the default width, item bounds, separator detection and separator-width fallback. They differed only in whether an embedded combo's width went through sibling negotiation.

Both now delegate to $toolbar_button_width_core with an explicit mode. The large-combo negotiation still calls the raw wrapper for siblings; calling the constrained wrapper there would recursively negotiate the same two controls. Existing entry points, sizing constants and lookup order are preserved. The WAT change removes 17 net lines.

test/test-toolbar-widths.js passes with both the original and refactored implementation. It covers missing records, nonpositive/default button widths, ordinary buttons, separator widths around both fallback boundaries, invalid indices, two large combo siblings, narrow/wide parents, and the small-combo exception. In a 400-pixel parent, raw widths 250/300 plus a 23-pixel sibling retain constrained widths 80/125; a wide parent retains 250/300. This tests the existing algorithm, not independent Win98 conformance of its 160-pixel threshold or 80-pixel minimum. Those heuristics were not introduced or broadened by this change.

The existing toolbar insertion/state/geometry suite passes 48/48, and WordPad's real-EXE toolbar layout/command regression passes 23/23. The control-variant gate's ToolbarState attribution moves from the two old bodies to the shared core; the wrappers no longer access fields themselves. The initial full build correctly rejected those stale attributions before they were updated. No gate rule was weakened. No performance benefit is claimed; this removes a duplicated sizing policy, not an interpreter hot-path optimization.

The corrected control-variant gate passes (579 sites, 200 attributed functions). A subsequent full build is not green: another active lane resized the DirectX regions in src/00-regions.wat while validation was running, and the build correctly rejected its stale lib/region-map.generated.js. This refactor changes no region declarations or map. The owner was notified on the messageboard; those files were not modified or staged here. The source-compiled width, toolbar and WordPad tests above completed before that concurrent resize. Logs are /private/tmp/wa-toolbar-width-build.log (initial attribution rejection), /private/tmp/wa-toolbar-width-build-final.log (map freshness rejection), /private/tmp/wa-toolbar-width-insert.log, and /private/tmp/wa-toolbar-width-wordpad.log.

Follow-up: the DX owner regenerated the map. The subsequent full build during the Win16 rectangle-bridge follow-up passed both canonical and compatibility artifacts, including this toolbar refactor and its corrected attribution. Log: /private/tmp/wa-win16-rect-build.log, layout da439e4cfa25a28d. The shared-build limitation above is resolved; the earlier failures remain recorded as such.