Migrate complex runes from Model to createContentModelSchema

Migrate the remaining runes that have custom processChildren overrides with stateful parsing, text pattern matching, or DOM surgery. All will use the custom content model type.

Runes

Rune	Location	Challenge
form	`packages/runes/src/tags/form.ts`	Multi-pass field parser with type inference from names, blockquote/list pattern detection, placeholder parsing, heading-to-fieldset conversion
conversation	`packages/runes/src/tags/conversation.ts`	Blockquote-to-message conversion, speaker name extraction from bold text, alternating alignment
tabs	`packages/runes/src/tags/tabs.ts`	Heading-to-tab conversion, image extraction from heading content, header/tab content separation
bento	`runes/marketing/src/tags/bento.ts`	Icon paragraph extraction via descendant tag filtering, size/span tier detection, DOM surgery separating icons from body
comparison	`runes/marketing/src/tags/comparison.ts`	Row type detection (check/cross/negative/text), bold label extraction, separator detection/removal, table structure mapping

Acceptance Criteria

form rewritten using createContentModelSchema with custom content model
conversation rewritten using createContentModelSchema with custom content model
tabs rewritten using createContentModelSchema with custom content model
bento rewritten using createContentModelSchema with custom content model
comparison rewritten using createContentModelSchema with custom content model
refrakt inspect <rune> --type=all output is identical before and after for each rune
All existing tests pass — especially form, tabs, and conversation tests which cover edge cases
No Model class import remains in any of the migrated files

Approach

Each rune's processChildren logic moves into the custom content model's processChildren function. The key difference is the return type — instead of setting instance properties via @group, the function returns a resolved fields object that the transform function receives.

Per-rune notes

form: The field parser state machine (type inference, blockquote detection, placeholder parsing) moves wholesale into processChildren. This is the largest single migration.
conversation: Speaker extraction from bold text and alternating alignment tracking move to processChildren. Relatively contained.
tabs: Heading content walking and image node extraction move to processChildren. The heading-to-tab tag reconstruction needs careful handling.
bento: The descendant tag filtering (icon paragraph detection) and body/icon separation move to processChildren. Size tier detection stays in transform.
comparison: Row type detection, label extraction, and separator removal move to processChildren. Table structure mapping stays in transform.

Dependencies

WORK-099 (simple core runes migrated)
WORK-101 (moderate runes migrated — validates the custom pattern on storyboard first)

References

SPEC-032 (parent spec)