De-barrel internal imports + lint/CI ratchet

[amcclain]

Part of #4640 - the mechanical repo-wide rewrite. Must land after #4643, verified by #4641's harness.

Goal

Internal code imports concrete modules; barrel index.ts files become public API surface only. This removes the amplifier that turns any two-module relationship into a whole-package cycle, and is what makes barrel elision (tree-shaking) safe to re-enable.

Sequencing constraint (important)

Today, internal barrel imports are load-bearing: inside a boot cycle the barrel is always already mid-evaluation, so re-entering it is a benign back-edge. Rewriting import {XH} from '@xh/hoist/core' to a deep import turns that back-edge into a forward edge into a not-yet-evaluated module - introducing the TDZ crashes this project exists to fix, even under sideEffects: false. The upward edges must be severed first (#4643), with the guard harness green, before this lands.

This is the mechanical, obvious-looking phase, which is exactly why someone will be tempted to take it first.

Scope (counts from develop, 88.0.0-SNAPSHOT)

Roughly 824 internal barrel imports to rewrite, against 1,941 deep module imports already in the target style - so this finishes a convention the codebase mostly follows rather than inventing one. Breakdown:

Import form	Count	Disposition
from '@xh/hoist/core'	562	rewrite
from '@xh/hoist/data', /format, /promise, /exception, /svc	229	rewrite
from './' / from '../' (own package barrel)	33	rewrite - includes both crash sites
from '@xh/hoist/icon'	213	keep - side-effectful barrel
from '@xh/hoist/mobx'	170	keep - side-effectful barrel

98 index.ts files in the package overall.

Work

• Codemod internal barrel imports to concrete module paths (ts-morph or similar).
• Preserve intentionally side-effectful barrels - the sideEffects array in package.json is the authority (icon/index.ts, mobx/index.ts, core/persist/index.ts, kit/blueprint/index.ts, the platform register.ts files, and the rest) - adding explicit side-effect imports where registration must still run.
• Land the ESLint guard (no-restricted-imports scoped to internal files) in the same PR, or IDE auto-import regresses it immediately. Note this sits alongside the foundation-set rules introduced in Sever upward eval-time edges: late-bind XH out of low-level modules #4643 and should be written as one coherent config, not two competing ones.
• Add cycle-detection ratchet to CI: existing cycles grandfathered, new ones fail.
• Coordinate timing against in-flight branches - this conflicts with everything open, and unlike Convert type-only imports/exports to import type / export type, with enforcement #4642 the conflicts are not regenerable by re-running a formatter. Wants a deliberately chosen quiet window and a fast merge, not a long-lived PR.

Done when

The #4641 harness is green across all barrels, the lint guard is on, and the CI ratchet holds the line.