fix(publisher): pay the KING curve the lane actually decided on

[ai-hpc]

The live publisher is not paying the decided curve. Found while clearing the held PRs.

What diverged

distill moved cybergym_tournament._award_shares to the KING model on 2026-09-04 (distill #181) — jared's spec: ranks 2–5 take the fixed runner-up shares (0.07 / 0.03 / 0.03 / 0.03) and the king takes the residual, so a thin field concentrates on the leader rather than scaling every rank up.

The publisher's vendored copy — the one the live weight path composes from — still had the old table and renormalized it:

winners	publisher (live)	distill (decided)
2	0.82 / 0.18	0.93 / 0.07
3	0.73 / 0.16 / 0.11	0.90 / 0.07 / 0.03
5	0.65 / 0.14 / 0.10 / 0.07 / 0.04	0.84 / 0.07 / 0.03 / 0.03 / 0.03

This PR re-syncs the vendored file from distill main (de916cf) and records that commit in the VENDORED header.

Why the drift wasn't caught

test_vendored_tournament_constants_match_the_mechanism pinned TOURNAMENT_SHARES. After #181 the two copies had different constants — RUNNER_UP_SHARES in distill, TOURNAMENT_SHARES in the vendored copy — rather than different values of one, so the guard kept passing against a name that no longer drove the payout.

test_vendored_award_shares_matches_distills_king_schedule now pins the schedule, rank by rank, for every field size, and asserts the lane pays out in full for any non-empty field. That is the thing that actually pays, and it is the same table distill's own golden test asserts.

Supersedes the held curve PRs

• feat(cybergym): v3 lane bottom-fill payout curve (leader takes remainder) #146 vendored an older bottom-fill table (0.65/0.14/0.10/0.07/0.04, leader takes remainder). Same shape, superseded values — closing it in favour of this.
• distill fix(independent): do not failover SAT policy refusals to another peer #161, its byte-identical twin, is likewise superseded: distill main already carries the KING curve, so merging fix(independent): do not failover SAT policy refusals to another peer #161 would have reverted the decided table.

wallscaler's review of #146 asked for the header sha to be bumped to the distill commit being mirrored; done here (de916cf).

Risk

The CyberGym lane currently has no winners to pay — the v2 backend went live today and no miner has submitted — so this is the cheapest possible moment to correct the curve. 52 passed in the adapter suite locally; the rest of the publisher suite needs bittensor, which CI has.

[ai-hpc]

Self-review before merge (owner-directed: jared asked me to clear the held PRs and deploy).

• Blast radius. The vendored file is imported only by mechanism_cybergym_adapter (epoch_score_base100, build_scoreboard, WINDOW, TournamentError); validator_thin never imports it, and Authority/FULL mode refuses v3 outright — so the curve is publisher-side at runtime. I checked every one of those names still exists on the synced file before committing.
• Direction of the change. It moves the live publisher ONTO jared's decided spec, not off it. The old behaviour was the drift.
• Timing. The lane has no winners right now (v2 backend went live today, no submissions), so no in-flight payout changes under anyone.
• The guard that failed. I did not just update the constant — the reason the drift survived a passing guard was that the two copies pinned different names. The new golden pins the schedule itself, which is what pays, for every field size, plus the sum-to-1 invariant.
• Not byte-identical to distill, by design: the six-line VENDORED header differs (it always has). Everything else is distill de916cf verbatim.

All checks green: Publisher suite, Python 3.11/3.12, Integration lane, Real validator release, AMD SEV-SNP contract.