fix(core): close isCompatible() correctness holes for cross-app duck typing

[cuibonobo]

Summary

Fixes #54. isCompatible() had two correctness holes and a spec example that contradicted the implementation, which matters because #47 promotes it to the official cross-app duck-typing mechanism.

• Optional-field hole: the predicate checked whether the candidate schema declared a field, not whether it declared it required. A candidate with an optional title passed a required-title check, so a consumer trusting isCompatible() could read undefined from a field it was told was guaranteed. Fixed by requiring field.required === true on the candidate.
• text vs string: these are distinct ScalarFieldKinds compared with strict equality, but the spec's own motivating example ({ text: string }) doesn't work under the shipped predicate since long-form text is typically declared kind: 'text'. Added a READ_COMPATIBLE table making string⇄text mutually acceptable (identical value sets) while every other kind pair stays exact — notably date is not compatible with string, since date carries a parse/validity guarantee a plain string doesn't.
• Shallow array/object matching: only top-level kind was checked, so a candidate's nested object.properties or array.items could silently drop a required sub-field. Now recurses using the same (strengthened) predicate.

Also scopes isCompatible() explicitly as read-compatibility (licenses consuming records, not writing them) in both the code doc comment and docs/spec.md §Type compatibility, whose inlined copy of the function is updated to match.

Test plan

• Added tests: optional-field-on-candidate now fails a required check; text/string accepted both directions; date/string still rejected both directions; nested object required-subfield mismatch; nested array item-kind mismatch (both a true mismatch and a text/string-equivalent pair); the spec's own example as a fixture.
• pnpm --filter @haverstack/core test — all 256 tests pass
• pnpm --filter @haverstack/core typecheck and lint — clean
• Confirmed via git stash that the pre-existing wire-types typecheck failure is unrelated to this change (missing built @haverstack/core output, not a type error)

---

Generated by Claude Code