feat(orb): publish per-arm precision, coverage, and Wilson intervals with the fleet accuracy

[JSONbored]

What

Every published accuracy figure now carries its per-arm split, the coverage it was earned at, its Wilson 95% interval, and the sample size behind it. Closes #8829 (epic #8828, Phase 1 — the measurement layer everything later phases depend on).

Why

Two structural problems with the bare scalar, both flagged in the epic's research basis:

1. It is gameable. Holds are correctly excluded from the accuracy denominator (they are deferrals, not decisions) — which means raising the hold rate mechanically raises published accuracy without any quality change. The counter is to publish the coverage alongside: "97% at 40% coverage" and "94% at 85% coverage" are different claims, and now they read differently.
2. It is unverifiable at small samples. At current volumes the interval around the point estimate is wider than the improvement the tuning roadmap is trying to demonstrate. Wilson intervals (never Wald, which claims impossible certainty exactly where a gate metric lives, near p→1) make the uncertainty visible: 4/5 confirmed renders as 80% with a ~36-96 interval, which is the honest statement.

Per-arm separation matters because the cost structure is asymmetric — a wrong merge costs more than a wrong close — so mergePrecision and closePrecision must be separately visible and separately CI'd ahead of the per-arm Neyman-Pearson thresholds (#8835).

How

• analytics: InstanceMetrics now carries the raw confusion counts — ratios can be neither pooled across instances (Simpson) nor intervalled. fleet.pooled sums them over eligible instances with coverage = verdicts / (verdicts + holds); policy actions (fix(orb): score enforcement closes as their own class, not quality mispredictions #8827) sit outside both sides. wilsonInterval() is the one shared, pure interval primitive, null on zero trials — no data must render as no claim.
• public-stats: fleetAccuracy gains mergePrecisionPct/CiPct, closePrecisionPct/CiPct, accuracyCiPct, coveragePct, decidedCount. Pooled and median views coincide exactly with today's single registered instance.
• UI: the hero hint appends "· at N% coverage"; the field is optional-chained so an older backend degrades to the bare figure.
• OpenAPI regenerated.

Verification

• TSC clean; 84 tests across public-stats/analytics/mcp-fleet/operator-dashboard plus 38 analytics tests and 10 UI tests, all green.
• New property tests pin the Wilson edges: 60/60 does not prove p=1, 0/20 does not prove p=0, more trials tighten, zero trials → null.
• Pooling test: multi-instance sum with holds and policy actions, unregistered instances excluded.
• Changed-line coverage: 0 uncovered statements/branches.

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.

[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	loopover-ui	745fbf2	Commit Preview URL Branch Preview URL	Jul 26 2026, 10:48 AM

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.29%. Comparing base (ee5e724) to head (745fbf2).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #8839      +/-   ##
==========================================
- Coverage   93.83%   92.29%   -1.55%     
==========================================
  Files         800      800              
  Lines       79851    79879      +28     
  Branches    24203    24209       +6     
==========================================
- Hits        74928    73723    -1205     
- Misses       3558     5092    +1534     
+ Partials     1365     1064     -301     

Flag	Coverage Δ
backend	92.97% <100.00%> (-2.15%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/openapi/schemas.ts	100.00% <ø> (ø)
src/orb/analytics.ts	100.00% <100.00%> (ø)
src/review/public-stats.ts	98.87% <100.00%> (+0.12%)	⬆️

... and 3 files with indirect coverage changes