feat(scores): add hybrid v2 score rollout

[Albermonte]

Summary

• add marker-gated, atomic activity synchronization and bounded repair
• persist score v2 alongside v1 with off, shadow, and active rollout modes
• expose current, stale, and no_score states without discarding trustworthy prior scores
• add a dashboard score-version selector with URL persistence across overview, validator list, and details
• infer missing high-stake election gaps as offline in v2 while preserving low-stake random non-election
• add score-sync progress logging, development full-repair support, migrations, and tests

Why

Partial activity history could make a score look complete when it was not, while short recent outages were heavily diluted by the long-term window. This keeps v1 available for rollback and comparison while v2 adds a recent-performance component and explicit data-freshness handling.

Missing election rows are ambiguous: a high-stake validator disappearing for several finalized epochs is strong offline evidence, but the same gap is normal for a low-stake validator. V2 therefore requires both at least 1% estimated network stake and less than 0.1% aggregate random non-election probability before marking a complete gap inferred_offline.

Representative score changes

Calculated with the actual computeScore and computeScoreV2 implementations using a representative 550-epoch long-term window, 28-epoch recent window, two epochs per day, and 720 blocks per epoch. Low stake is modeled as 1% network dominance and high stake as 10%. Values are rounded synthetic examples, not production forecasts.

Case	Score v1	Score v2	v2 − v1	Important behavior
Healthy, low stake	100.0%	100.0%	0.0 pp	Healthy baseline stays unchanged.
Offline for 1 day, low stake	98.3%	96.4%	-1.9 pp	Recent downtime becomes visible sooner.
Offline for 3 days, low stake	95.4%	89.3%	-6.1 pp	A multi-day outage is no longer diluted by months of history.
Offline for 14 days, low stake	83.6%	49.8%	-33.8 pp	Sustained recent unavailability has a strong effect.
Healthy, high stake (10%)	95.2%	95.2%	0.0 pp	The existing dominance penalty remains shared by both versions.
Offline for 3 days, high stake (10%)	90.8%	85.0%	-5.8 pp	Dominance and recent outage penalties compound.
Healthy at the 15% dominance threshold	0.0%	0.0%	0.0 pp	The shared dominance component reaches zero at the configured threshold.
Miss 7 of 720 blocks every epoch	94.1%	94.1%	0.0 pp	Persistent degraded performance remains primarily a reliability signal.
Miss 50% of blocks for the latest 3 days	97.5%	89.3%	-8.3 pp	v2 reacts to a sharp recent reliability regression.
Voluntarily inactive for the latest 3 days	100.0%	89.3%	-10.7 pp	v2 treats deliberate inactivity as lost recent availability.
Not elected by protocol randomness for 3 days	100.0%	100.0%	0.0 pp	Random non-election remains neutral rather than being treated as downtime.

Missing-election guardrails

Finalized activity gap	V2 classification	Score impact
Estimated stake below 1%	not_elected_randomness	Neutral
Estimated stake at least 1%, but random non-election probability at least 0.1%	not_elected_randomness	Neutral
Estimated stake at least 1% and random non-election probability below 0.1%	inferred_offline	Reduces v2 availability
Incomplete epoch coverage	No current v2 score	Keeps prior trustworthy score stale when available

Rollout and safety

• off: compute and serve v1
• shadow: compute both versions and continue serving v1 by default
• active: compute both versions and serve v2 by default
• API and dashboard callers can explicitly request score-version=1|2
• current v2 is withheld when recent activity coverage is incomplete
• partial long-term coverage can retain the last trustworthy v2 as stale
• production repair remains bounded; development mode supports a full repair and a shorter orphaned-sync lease

Verification

• pnpm test run — 193 tests passed
• pnpm typecheck — passed with Node 24
• pnpm lint — passed
• pnpm validate:json-files — passed
• pnpm build — passed with Node 24
• git diff --check — passed
• local sync:scores — complete at epoch 1254 with 100% recent and long-term coverage; 53 v1 and 36 v2 scores

Targeted local verification confirms a validator observed with one slot (0.195% stake estimate) keeps 231 missing epochs as not_elected_randomness, while a validator near 6.4% stake retains its 20 high-confidence inferred_offline epochs.

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	validators-api-test	9d44e04	Commit Preview URL Branch Preview URL	Jul 31 2026, 03:28 PM

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	validators-api-main	9d44e04	Commit Preview URL Branch Preview URL	Jul 31 2026, 03:27 PM