refactor: collapse the skill predicates and give the reroll a home

[philoserf]

Closes #50, #44, #43, #39 — all of milestone 5, one PR, since all four land in the same ~40 lines.

One predicate instead of two (#50, #44)

canSpendSkill and canSpendSkillButton were the same lines with === flipped to !==, called one after the other in the same render. There's one underlying fact — this skill applies to this attribute and hasn't been spent — and one variable on top of it: has the player already pressed the button this paragraph. skillAvailableFor is the fact; the call site derives the rest.

makeSkillButton inlines — eight lines with one caller that always passed () => v.repaint(), so the parameter never varied.

That collapse is what deletes #44. if (skillBonusActive) v.state.draft.skillUsedHere = attr; could never change anything, because skillBonusActive already required skillUsedHere === attr. It's a symptom of the split rather than a defect of its own, so it disappears with the restructure rather than as a separate edit — which is what both issues predicted.

The reroll gets a home and its first test (#43)

planRoll set rerollPolicy: 'highest' and stopped; the mechanic lived 400 lines away inside the roll button's callback. Neither file answered the question a maintainer actually has — rules.ts didn't say what a reroll costs, play.ts didn't say which scenarios trigger it.

applyReroll(dice, policy, rng?) now sits in rules.ts beside the RollPlan that requests it, and the roll callback drops to two lines. roll already took an injectable RNG, so it's deterministically testable — and nothing tested the splice before: tests/rules.test.ts had two reroll_highest cases and both asserted on the flag.

Six cases added, including the one that matters:

test('replaces a highest die that was already a success', () => {
  const after = applyReroll([6, 2], 'highest', always(1));
  expect(after).toEqual([1, 2]);
  expect(countSuccesses(after)).toBe(0);
});

That's what makes reroll_highest a hazard rather than a bonus, and it was precisely the half that couldn't be tested where the code sat. Probed by making the reroll "helpfully" spare a successful die — three tests fail.

#39 finishes

The hand-written 'penmanship' | 'language' | 'heart' union is gone from the file; Attribute is imported. Two of its five original sites had already gone — one with the draft's move to src/paragraph.ts, one with #46 — which is why this lands as a two-line change rather than five.

Verification

bun run check:ci clean. 77 → 83 tests. src/screens/play.ts 695 → 677.

Smoke-tested on The Art Dealer — the only scenario carrying reroll_highest — with Illumination, so the skill bonus and the reroll land on the same Penmanship roll:

Step	Info line
Penmanship step	Roll Penmanship (3 dice, re-roll the highest) for a fine hand.
after spending the skill	Roll Penmanship (4 dice, re-roll the highest, skill applied) …
paragraph II	Roll Penmanship (3 dice, re-roll the highest) …

Button disappears on spend, reference card reads The Monk — Illumination spent, and no button reappears on paragraph II. That's the check #50 asked for, plus the interaction with the reroll.

Note

Two of these issues were smaller than filed, because milestone 4 had already taken part of them: #50's complaint about a triple skillById lookup was resolved by #46, and #39 was down from five sites to two. What remained was the genuine twin collapse and the mechanical import.

THEORY.md and WALKTHROUGH.md untouched.