refactor(core): move the Inbox thread's pure rules into @plannotator/core

docs/agent-guides/inbox-compose.md

@@ -52,7 +52,7 @@
 - The catalog: `GET /api/inbox/skills` (`packages/server/inbox-skills.ts`), plan review's `handleReferenceSkills`: the global skill folders only, read fresh. The window installs it as ui's `skillCatalogTransport`, so the comment composers on attached files list it too, and sets `skillInstructionsDestination: 'your reply'` (their human-only line says the instructions go with the reply).
 - At Send the window composes ONE block at the end of the message (`packages/inbox/skills.ts`, Option B: the client composes the body, as plan review composes its feedback): references are read from the words and every comment on an attached file the Send carries (not a guided review's diff comments), once each; `renderSkillReferencesBlock` (ui) writes Plannotator's block in "the person … this reply / this message" wording (`@plannotator/core/skill-reference-block`): a model-invocable skill by name, a human-only skill's SKILL.md (read then from `GET /api/inbox/skills/content`, 20,000-character cap, truncation notice, marker lines neutralized) between BEGIN/END markers with `Skill directory:` naming its folder. At most 40,000 characters of instructions per message (`INBOX_SKILL_INSTRUCTIONS_BUDGET`); past it, or unreadable, a skill is named with "read SKILL.md in <dir>". A reply carries the block in its `feedback` (after the annotations' feedback, so it is the body's end; the server is unchanged); a New message appends it to `body`. The per-comment "Skills referenced" list is off in the Inbox (`configurePlannotatorUI({ skillReferenceExport: false })`), so a comment's `$name` stays plain text and the block covers it.
 - Trust: only the person's Send composes a block; MCP has no skills field, and an agent's `$readback` is plain text. The window folds a block only under the person's own messages: `MessageBody` splits it off with `splitInboxSkillsBlock`, which takes the LAST Inbox heading (after a blank line) whose tail is exactly the shape the window composes: list lines, then nothing but blank lines and closed BEGIN/END sections for skills the list includes, to the end of the body. The real block is always last, so agent text a reply quotes (a guided-review comment's diff lines, an attached-file comment's `Feedback on:` quote) can never fold the person's feedback, not even with a fake unterminated BEGIN running on to the real END; fence parity is not read, since the agent controls it. Both code exporters (the review app's `exportFeedback.ts` and the Inbox's `guideCommentEntry`) also fence quoted code with more backticks than any run in it (`codeFenceFor`, `@plannotator/core/markdown-structure`; unchanged bytes for code without a backtick run). It draws `SkillsLine` (`components/SkillsLine.tsx`): "Skills: readback, write-better · readback's instructions went with this reply (0.4 KB) · Show", a `<details>` that opens on the block as the agent got it. The references in the sent words are painted as the composer painted them (`skillHighlight.ts`: the CSS Custom Highlight API, `::highlight(ib-skill-ref)` / `ib-skill-ref-human` with the dotted underline, matched against the skills the block names, never the live catalog, only in the person's own words (before the answers or feedback headings), at a priority below search's landing highlight; nothing painted where the API is missing). An agent's message is shown whole.
-- Snippets and search read the person's words: `replyWords` and the folded line's snippet (`folding.ts`), and search's `inboxSearchReplyText`, drop the block (the `$name` stays in the words, so search still finds the reply by the skill's name; the SKILL.md text is not indexed for that reply). The stored body, `read_thread`, `wait_for_reply`, the wake and the phone carry the whole text (the accepted cost of Option B).
+- Snippets and search read the person's words: `replyWords` and the folded line's snippet (`@plannotator/core/inbox-folding`), and search's `inboxSearchReplyText`, drop the block (the `$name` stays in the words, so search still finds the reply by the skill's name; the SKILL.md text is not indexed for that reply). The stored body, `read_thread`, `wait_for_reply`, the wake and the phone carry the whole text (the accepted cost of Option B).
 - Phone (`PHONE_QUERY`): the picker opens full width directly above the reply box (no sheet, so the keyboard stays up and a tap inserts on pointerdown), 44 px rows with the description under the name, at most about 4½ rows.
 - On the phone's device door (contract 7.46, 7.47). Tests: `packages/server/inbox-skills.test.ts`, `packages/inbox/skills.test.ts`, the DOM test `packages/inbox/skillReferences.test.tsx` (the whole window over a real Inbox; the DOM server pins the skill folders under the test's scratch HOME).