feat(evals): size-appropriate werewolf role balance + private non-voter re-prompt

[Poytr1]

What

Stacked on #984/#985/#986 (retargets as they merge). Fixes the composition flaw the user caught — two wolves at a five-player table degenerate: one unsaved night-1 kill reaches parity instantly, so every 5p game was a one-night game with no day, vote, or later round (measured across all real 5p runs).

Role table (werewolfWolfCount, shared by both game compositions):

players	werewolves
5–6	1
7+	2

Single-wolf night mediation collapses to propose-only — the lone wolf's clear statement IS the pack's choice (referee call text + role message adjusted; the absorb path already took a partnerless proposal directly). The two-wolf mediated propose→relay→agree leg stays pinned by a new 7p scripted cell (three rounds, every kill verdict: 'agreed').

Non-voter re-prompt (the human-moderator mirror, now on the critical path since votes decide 5p games): when a vote goes quiet the host nudges the referee — the first nudge sends each living non-voter ONE private needsReply re-prompt mandating the self-identifying ballot form (direct replies carry no sender label); the second closes the ballot counting the still-missing as abstentions. Private ballots parse from direct, coalesced, and #800-inferred replies alike; either channel settles the obligation. Pinned by a new abstainer cell: a scripted public-vote abstainer is re-prompted once, its private ballot counts, zero abstentions, game completes.

Also fixes a deterministic scripted-player bug the new multi-day games exposed: the previous round's VOTE rows lag into the next day's prompts and were counted as SPEECH, convincing seat 4 the whole round had already spoken — vote lines are votes, never discussion speech.

Sanctioned re-pins (the game-design change): 5p seed-1 is now a two-round village win through actual play (night-1 save → full day-1 ballot → seer inspects the wolf → day-2 lynch); 6p keeps one wolf. Slack-shaped werewolf suite passes unmodified (its pinned cells are 7+ players, unaffected by the 5–6 row).

eval:collab:contracts now 122 tests; eval:parity 23, eval:contracts 46, evals tsc — all green.

🤖 Generated with Claude Code

[agentconnect-md-test]

Approved. I found no blocking regressions in the exact reviewed revision.

The size-based role table is shared cleanly by both werewolf compositions, the single-wolf path correctly bypasses the partner-verdict leg, and the vote nudge state machine gives missing voters one private collection pass before closing remaining absences as abstentions. I also checked the multi-round scripted-player change against the existing lagged vote-row behavior.

Verification completed:

• eval collaboration contracts: 122/122 passed
• eval parity: 23/23 passed
• eval contracts: 46/46 passed
• focused webchat + Slack-shaped werewolf tests: 23/23 passed
• eval TypeScript check passed
• exact diff whitespace/conflict check passed

_{sent by review-bot (Codex · gpt-5.6-sol) · open in session}