chore(agents): teach parity + review loops idiomatic Ember patterns

[patricklx]

What

Gives the automated parity-fix agent (and its review+fix loop) real guidance on writing idiomatic Ember, instead of transliterating the Carbon React source.

Why

The agent's instructions amounted to "Simplify: don't overcomplicate React patterns". That's not enough signal to pick between, say, yielding a contextual component and iterating over what was passed in — so generated components tended to mirror the .tsx structure. The review+fix loop had no idiom dimension at all, so un-Ember-ish structure survived all 5 rounds.

Changes

AGENTS.md — new "Idiomatic Ember Patterns" section:

• A React → Ember translation table (React.Children.map/cloneElement, component-as-prop, controlled/uncontrolled pairs, useRef+useEffect, forwardRef, createPortal, context, useId, Floating UI, debounce, effect cleanup, lazy async values).
• Worked examples for the six that need more than a table row.
• A "What NOT to Reach For" list: @ember/render-modifiers, A()/pushObject, set() for tracked state, classic Component+.hbs, this.element.

Every entry points at a component already in this repo — tabs.gts, tree-view.gts, data-table.gts, link.gts, text-area.gts, -private/tooltip.gts, portal.gts, search.gts, services/*, the generated icons/*.ts. Nothing here is aspirational.

The "Simplification Guidelines" section is reframed: simple means fewer moving parts for the consumer, not fewer lines in the component.

.github/workflows/templates/agent-prompt.md

• New required Phase 1 step 4: map each React construct to its Ember idiom before writing code, with an abbreviated table inline.
• Both implementation paths (A: fix existing, B: new component) now build on that mapping, with an explicit scope limit — modernise legacy patterns in code you're already touching, don't sweep the repo.
• Three new success criteria (idiom translation, no new "What NOT to Reach For" usage, fully typed signature).
• Fixes a wrong note: "Native Helpers: element is built-in, don't import it". It isn't — it comes from ember-element-helper, which AGENTS.md's Pitfall 1 has said all along.

scripts/fix-parity-issue.sh

• The review prompt gets a concrete list of un-Ember-ish tells to check, and must name the replacement pattern plus an existing component using it, so findings are actionable. Scoped to what the branch touches; explicitly told not to invent nits to justify another round.
• The fix prompt asks for the actual restructure, and to justify any finding it declines.
• The PR-review and fallback prompts reference the same section.

Verification

• bash -n clean; shellcheck shows only the 4 pre-existing informational findings.
• Template renders correctly through the script's sed substitution chain — no placeholder collisions (the {{#in-element}} in the new table is untouched by every substitution rule).
• Cross-checked each documented pattern against the source it cites, and corrected two drafting errors in the process: TrackedPromise lives in the generated icons/*.ts, not icon.gts, and <Tabs> yields a single positional Pane ({{yield (component TabPane tab=this)}}), not a namespace.

Docs/tooling only — no runtime code touched.