feat: multi-attestation aggregation — consensus verdict before TrustScoreJob

[mdproctor]

Context

When multiple agents attest to the same ledger entry (two SOUND, one FLAGGED), TrustScoreJob currently treats each attestation independently — three separate Beta updates. This creates noise: the FLAGGED attestation from a less reliable peer carries the same weight as two SOUND attestations from highly trusted peers. Aggregating attestations to a consensus verdict before scoring reduces noise.

Part of: #49 (Group A epic)

What to implement

New component AttestationAggregator in the trust scoring pipeline:

@ApplicationScoped
public class AttestationAggregator {

    /**
     * Given multiple attestations on the same ledger entry,
     * produce a single consensus verdict and confidence.
     *
     * Strategy: confidence-weighted majority.
     * - Sum confidence-weighted SOUND/ENDORSED votes
     * - Sum confidence-weighted FLAGGED/CHALLENGED votes
     * - Consensus = majority verdict, confidence = margin / total
     */
    public Optional<AggregatedAttestation> aggregate(List<LedgerAttestation> attestations) { ... }

    public record AggregatedAttestation(
        AttestationVerdict consensusVerdict,
        double aggregatedConfidence,
        int attestorCount
    ) {}
}

Update TrustScoreJob to:

1. Group attestations by ledgerEntryId
2. For entries with multiple attestations, call AttestationAggregator.aggregate()
3. Use the aggregated result as a single trust update rather than N individual updates

Aggregation strategy (configurable):

• WEIGHTED_MAJORITY (default) — confidence-weighted vote
• UNANIMOUS_REQUIRED — only SOUND if all attestors agree
• FIRST_ATTESTOR — use first attestation only (legacy behaviour)

casehub.ledger.trust.aggregation-strategy=WEIGHTED_MAJORITY

Tests

• 2 SOUND (confidence 0.8) + 1 FLAGGED (confidence 0.3) → consensus SOUND, aggregated confidence reflects margin
• 1 SOUND (confidence 0.5) + 1 FLAGGED (confidence 0.9) → consensus FLAGGED (higher total weight)
• Single attestation → aggregated result equals that attestation unchanged
• UNANIMOUS_REQUIRED strategy: any FLAGGED → consensus FLAGGED
• FIRST_ATTESTOR strategy: identical to current behaviour (regression)

Motivating use case

A PR is reviewed by three agents. Two highly-trusted reviewers (confidence 0.9 each) mark it SOUND. One lower-trust reviewer (confidence 0.3) marks it FLAGGED due to an overly conservative style concern. Weighted majority: SOUND wins with strong margin. The assigning agent's trust score increases rather than being dragged down by the minority FLAGGED.

Acceptance criteria

• AttestationAggregator CDI bean implemented with weighted majority as default
• TrustScoreJob updated to aggregate before scoring
• Three aggregation strategies implemented and config-driven
• At FIRST_ATTESTOR strategy, all existing trust score tests pass unchanged (regression)
• New tests covering each strategy and edge cases

Refs #49