docs: harden RollPart spec from design review #114

[edloidas]

Landed the 10 spec corrections to .claude/docs/STAGE3.md §5 from the design review of #84. Docs-only — no source or test changes. On merge, #84 is unblocked.

• Added total: number to literal and variable so result.parts.total === result.total holds for trivial expressions.
• Added the grouped variant — keeps user-typed (...) distinct in the part tree and preserves the 1:1 AST↔RollPart mapping (16 variants).
• Replaced opaque modifier: string with ModifierSpec[] mirroring flattenModifierChain; chains like 4d6kh3kl1 round-trip per-spec.
• Preserved threshold conditions on reroll (condition), explode (threshold?), successCount (threshold + failThreshold?), and critThreshold (successThresholds[] + failThresholds[]).
• Added count: number to dice — symmetric with fateDice, eliminates the "derive from rolls.length" footgun under keep/drop.
• Pinned successCount.total = successes − failures as an invariant regardless of empty-pool or wrapping arithmetic.
• Pinned group.keptIndices semantics — 0-based, indexes parts[] in author-written order, pre-sort, omitted on bare groups, written by evalGroupModifier onto the inner group part; dropped sub-rolls retain their full RollPart including versus.degree.
• Documented DieResult reference sharing between part-tree rolls[] and result.rolls[] — no deep-clone.
• Added the RollPartType = RollPart['type'] alias for exhaustive switches.
• Added per-branch construction notes for evalDice, evalModifier, evalGroupModifier, evalGroup, evalVersus, evalSuccessCount, and evalCritThreshold; rejected the partsStack alternative and the optional parts? field in prose.

Closes #114

_{Drafted with AI assistance}