docs: bring every package's AI.md back in line with what it publishes

[amritk]

The shipped AI.md files had gone stale in the way generated-and-committed
docs always do: silently, and only for the audience that cannot file an
issue about it. CI regenerates llms.txt and diffs it, which catches a stale
bundle — but a stale source regenerates perfectly, so an AI.md that never
mentions the subpath a release added stays wrong forever.

What had drifted:

• @amritk/mini documented neither watch nor template, nor the typed
  matchRoute / buildPath re-exports on /router, nor Field on
  /forms, nor the /vite subpath at all.
• @amritk/mini-lynx was missing computed / effectScope,
  fadeTransition, keepAboveKeyboard / focusedField and
  HANDLER_PREFIX.
• @amritk/lynx-notifications documented neither its /testing subpath nor
  the fake behind it, and had no Status section.
• All four native packages exported MODULE and EVENTS with nothing
  saying what they are for, and only lynx-dialogs showed how to wire a
  fake into installNativeBridge — the one thing a consumer testing its own
  screens needs.

Two accuracy fixes matter more than the additions. Every native package's
Status section claimed the Objective-C compiles against the real Lynx pod;
the macOS job was disabled on cost, so it now compiles only when somebody
runs pod lib lint by hand, and the docs say so. And lynx-dialogs carried
no Status section at all, so nothing in it told a reader that none of it has
run on a device.

bun run check:ai-docs is the guard: it reads each package's exports and
fails on a runtime export, a published subpath, or (for a package shipping
native sources) a Status section the AI.md never mentions. It runs early in
CI, before the build, because it needs neither. Exports no consumer ever
writes — the tree operations the JSX transform calls, and the like — are
listed in INTERNAL_EXPORTS with the reason, so a new export cannot join them
by accident.

Types are deliberately out of scope: an AI.md documents shapes by writing
them out, and requiring the identifier would push these files toward being a
second copy of the .d.ts, which is the one thing they must not become.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com

📦 Bundle-size delta vs main (5ce0ea4 → 530d7d6)

Package	Entry	main	PR	Δ
mini	core (.)	3,156 B	3,156 B	+0 B (+0.0%) ⚪
mini	flow	4,083 B	4,083 B	+0 B (+0.0%) ⚪
mini	router	4,259 B	4,259 B	+0 B (+0.0%) ⚪
mini	forms	4,259 B	4,259 B	+0 B (+0.0%) ⚪
mini	query	1,988 B	1,988 B	+0 B (+0.0%) ⚪
mini-lynx	core (.)	5,605 B	5,605 B	+0 B (+0.0%) ⚪
mini-lynx	flow	4,981 B	4,981 B	+0 B (+0.0%) ⚪
mini-lynx	router	6,130 B	6,130 B	+0 B (+0.0%) ⚪
mini-lynx	forms	5,400 B	5,400 B	+0 B (+0.0%) ⚪
mini-lynx	query	1,990 B	1,990 B	+0 B (+0.0%) ⚪
mini-lynx	testing	2,464 B	2,464 B	+0 B (+0.0%) ⚪

_{Gzipped bytes of each bundled entry (esbuild, minified, browser/es2022, optional peer deps external). Bundling is deterministic, so unlike a timed benchmark these numbers carry no noise and every Δ is exact. ⚪ within ±0.5% · 🟢 smaller · 🔴 larger. Each package's core (.) must stay flat as subpath features land — that is the whole charter, and src/core-size-budget.test.ts holds the absolute ceiling.}