[Rust port] relayburn-analyze: findings + quality

[willwashburn]

Parent: #244
Depends on: #242, #266 (analyze foundation)

Context

The shared WasteFinding model and per-session quality outcome inference. Both are leaf modules — every hotspot detector and the patterns module emit WasteFindings through this surface, and compare / hotspots consume QualityResult to gate severity. They land together so the downstream detector PRs (#TODO ghost-surface, #TODO tool-call-patterns, #TODO patterns) can depend on a settled finding type.

Scope

findings.rs (port of findings.ts, ~313 LoC)

• Port WasteSeverity (info | warn | high), WasteAction (string-literal union — model as a string-newtype or a closed enum; either works, but emit identical strings at serialization), EstimatedSavings, WasteFinding.
• Port the *ToFinding adapters: retryLoopToFinding, failureRunToFinding, cancellationRunToFinding, compactionLossToFinding, editRevertToFinding, editHeavyToFinding, skillRecallDupToFinding, skillPruningProtectionToFinding, systemPromptTaxToFinding.
• Port findingsFromPatterns(result) and sortFindings(findings).
• The pattern types referenced (RetryLoop, FailureRun, etc.) live in patterns.ts (separate sub-issue). To avoid a circular-dep blocker, define those concrete struct types in findings.rs (or a shared patterns_types module) and have the patterns sub-issue depend on this one. Either layout is acceptable as long as findings doesn't depend on patterns (which is the larger, later module).

quality.rs (port of quality.ts, ~291 LoC)

• Port OutcomeLabel (completed | abandoned | errored | unknown), OutcomeConfidence (high | medium | low), SessionOutcome, OneShotMetrics, QualityResult, ComputeQualityOptions.
• Port inferOutcome(), computeOneShotRate(), computeQuality().

Conformance gate

• packages/analyze/src/findings.test.ts (322 lines)
• packages/analyze/src/quality.test.ts (259 lines)

Equivalent Rust tests; severity/action labels and outcome labels emit byte-identical strings.

Acceptance

• cargo test -p relayburn-analyze green for findings + quality.
• Public surface: WasteFinding, WasteSeverity, WasteAction, EstimatedSavings, all *ToFinding adapters, findingsFromPatterns, sortFindings, OutcomeLabel, OutcomeConfidence, SessionOutcome, OneShotMetrics, QualityResult, ComputeQualityOptions, inferOutcome, computeOneShotRate, computeQuality.
• sortFindings ordering matches TS — by (severity desc, savings desc, kind, sessionId, ts) tuple. Include a fixture-driven test.
• Pattern struct definitions placed so the patterns sub-issue can import without a circular dep.

Files

• packages/analyze/src/findings.ts + findings.test.ts
• packages/analyze/src/quality.ts + quality.test.ts