refactor: move the draft into a closure and widen the screen contexts

[philoserf]

Closes #42 and #46. Second and last PR of milestone 4.

The play screen owns its own state (#42, play half)

currentDraft, scenarioRecallOpen and lastSessionId were module-level lets — the app's only mutable state outside the store — because every renderPlay call was a fresh invocation that had to find the draft where the last one left it. renderPlay now creates that state itself and repaints its own subtree.

Four pieces of machinery existed only to work around that, and all four are gone:

• ensureDraftFor and lastSessionId, which detected a session change by hand. A new letter is now a new renderPlay call, so there is nothing to detect.
• The reset-before-onUpdate ordering, whose own comment warned that getting it backwards stranded the player on the done screen until they reloaded. Committing discards the whole closure — there is no ordering left to sequence, so the comment is deleted rather than preserved.
• PlayCtx.repaint, which pointed at main.ts's full-tree render. The two channels are now named apart: main.ts's commit() is durable, the play screen's internal repaint() is volatile.
• onUpdate — a general "mutate the session however you like" hook — narrowed to onCommit(paragraph). Committing is the screen's only store write.

The trap, handled deliberately. The render helpers take a per-invocation view, and state is one shared object per letter rather than a value copied into each render. That matters for exactly one thing: the roll button's 250ms shake timer reads v.state.draft when it fires, not a copy captured when its button was built. Capturing it would mean a repaint mid-shake lands the roll on a draft nothing will ever see — which is the bug the module-level let was accidentally avoiding.

Screens stop re-resolving what main.ts proved (#46)

PlayCtx and ScoreCtx carry the resolved character and skill. Seven characterById/skillById calls inside src/screens/ go, along with the fallbacks that invented five different answers to a question already answered: 'skill', '', 'The Correspondent', an internalError panel, and a thrown Error. toMarkdown takes the resolved pair instead of the whole CHARACTERS/SKILLS arrays, dropping its own find calls and its throw.

Verification

bun run check:ci clean, 77 tests. None of this has automated coverage, so it was driven in the browser:

Check	Result
Recall toggle mid-step	roll button is the same node — toggles in place, no repaint
Spend the skill	repaints, button disappears, info reads Roll Heart (5 dice, skill applied)
Every phase transition	0 localStorage writes
Commit	exactly 1 write
After commit	"Dip your quill…" / Paragraph II of V — a fresh draft, not a re-shown PARAGRAPH_DONE
Reload mid-letter	committed paragraph restored, Augmentation spent still latched, fresh draft at PICK_WORD
Widened context	reference card reads The Courtier — Augmentation unspent from v.character/v.skill

That fifth row is the one worth naming: it's the failure the deleted comment warned about, and it's now structurally impossible rather than avoided by statement order.

Also

Two comments corrected while in the file — attachRollButton named the recall toggle as a repaint cause, which stopped being true when that toggle became an in-place mutation; and the recall toggle's own comment still claimed a repaint would rewrite localStorage, which stopped being true in milestone 3.

src/screens/play.ts 709 → 695 lines, no module-level state left.

THEORY.md and WALKTHROUGH.md untouched, per the standing rule.

Merging closes milestone 4.