Unify one-agent contradiction policy: latest-wins vs both-count vs withhold across per-card and group tallies

[WilfordGrimley]

Follow-up from PR #482's Tron delta gate (2026-07-25) — deliberately NOT decided inside that fix cycle.

The codebase has no single contradiction policy for one agent holding conflicting printing votes on the same target:

• Humans, per-card: LATEST WINS (views.py delete-then-create on (card, anonymous_id)) — a human cannot hold two conflicting votes on one card.
• Machines, per-card: BOTH COUNT (bulk_create(ignore_conflicts=True) against the (card, printing, anonymous_id) constraint) — one machine agent CAN hold X and Y on a single card, and today both enter the tally.
• Any agent, group-level (post-Pool printing-consensus votes across md5 identity groups #482): WITHHELD entirely (option (iii) — contradiction across byte-identical members contributes nothing for the group).

Decide the unified policy. Tron's enumeration: withhold-vs-latest-wins for humans diverges only in the narrow case (h1: X on sibling A then Y on sibling B; h2: Y) — withhold gives None, latest-wins gives Y resolved; withholding never gains a wrong resolution, only loses recall. The machine both-count half is arguably a live per-card defect independent of groups (a self-contradicting machine agent inflates the denominator on one card today). A latest-wins-for-humans change is a pure function change in pool_group_votes + a recompute; the named test test_self_contradicting_human_withheld_not_latest_wins (landing on #482) must be consciously flipped by whichever ruling lands.

Requires an owner ruling + theory.md touch; PROTECTED CORE adjacency; do not build without the design conversation.

[WilfordGrimley]

Owner design conversation 2026-07-28/29 — withhold considered, NOT ruled; and why the WTC illustration surface does not prejudge this

Withhold is already this codebase's stated philosophy, just not applied uniformly

vote_consensus.py:163 implements withhold at the group level and grounds it explicitly:

Self-contradiction withholds. An agent whose votes across the group argue for different outcomes contradicted itself about byte-identical bytes, so it is not evidence for either outcome. This mirrors g₄'s withhold-never-manufacture philosophy (docs/theory.md).

So "null them both" is not a new option — it is the house principle, applied in one place out of three.

Adopting withhold uniformly would also close #525's machine half

The three live policies restated:

agent	scope	policy today	mechanism
humans	per-card	latest wins	views.py delete-then-create on (card, anonymous_id)
machines	per-card	both count	bulk_create(ignore_conflicts=True) vs (card, printing, anonymous_id)
any	group	withheld	post-#482 pool_group_votes

The machine row is exactly the defect in #525 — one calculator holding votes for mutually exclusive printings, all entering the tally at full machine weight. A uniform withhold ruling would fix that as a consequence rather than as a separate patch.

Blocker for the human per-card half: the contradiction is currently unrepresentable

Withhold requires KEEPING both votes so the tally can discard them. Today a human physically cannot hold two conflicting votes on one card — the submit view deletes any prior vote for (card, anonymous_id) before creating. So switching humans to withhold is not only a pool_group_votes change; it requires removing delete-then-create AND revisiting the uniqueness constraints, since the row that would record the contradiction cannot exist.

This is more than the "pure function change + recompute" this issue's original text anticipated for the humans half. Worth sizing before ruling.

Also unchanged from the original: test_self_contradicting_human_withheld_not_latest_wins must be CONSCIOUSLY flipped by whichever ruling lands.

The WTC illustration surface (#503) deliberately does not answer this

Phase C derives a CardArtistVote from a user's illustration pick (illustration_id -> artist is functional). That raises the obvious collision: the voter may already have an explicit artist vote on the card.

Ruling for that surface: a derived vote never overrides an explicit one.

This is an EVIDENCE-PRECEDENCE rule, not a contradiction policy. The voter typed an artist name; they merely IMPLIED one by picking artwork. An implication does not overwrite a statement.

It is forward-compatible with any ruling here, because withhold/latest-wins/both-count only engage when two votes of EQUAL STANDING coexist — and this rule prevents that state from arising in the first place. Whichever way this issue lands, phase C needs no revision.

Recorded so a future reader does not mistake phase C's behaviour for an implicit ruling on this issue. It is not one.