feat(export): add thinking_ratio to daily and session summaries (#633)

[lis186]

Summary

• Adds thinking_ratio (FLOAT, 0–1) to both daily and session export rows
• Ratio of thinkingDuration / elapsed across proxy-observed turns only
• null when no proxy turns with elapsed data exist (imported-only days/sessions)
• Not a timestamp or elapsed timing — a dimensionless ratio; does not touch INV-2

Why this field

Adversarial review by two independent reviewers (Fable 5.1 + Sol) confirmed:

• first/last_received_at timestamps → blocker (INV-2 violation) → dropped
• total_elapsed_s / total_thinking_s → major (INV-2 "elapsed timings") → dropped
• thinking_ratio → not an issue (dimensionless ratio, not a timing)
• Session duration available from Anthropic telemetry (duration_ms) → don't duplicate

Thinking ratio is the one metric only ccxray can provide: "what fraction of wait time was extended thinking" is invisible to Anthropic's per-request telemetry.

Implementation

• parseFloat(entry.elapsed) guards added (field is a string from toFixed(1))
• Daily/session accumulators: _totalThinkingS, _totalElapsedS, _elapsedTurns
• Ratio rounded to 4 decimal places, computed in finishDaily/finishSession
• Test golden normalizer updated; schema additivity test updated

Second review — codex gate

Reviewer: codex (gpt-5.6-terra), 4 rounds.

Round	Findings	Disposition
R1	2 major: elapsed not guarded on !imported; elapsed missing from mergeEntry fill-if-null	Fixed
R2	2 major: thinkingDuration also needs !imported guard; receivedAt/elapsed paired merge mixes copies	Fixed
R3	1 major: else branch still mixes timing pair	Fixed (simplified to atomic-only)
R4	1 minor: ratio can exceed 1 due to rounding mismatch	Fixed (Math.min clamp)

codex gate clean

Test plan

• 68 existing export tests pass (golden baselines, schema additivity, dedup, cross-agent)
• Imported-only entries produce thinking_ratio: null (no false zero)
• parseFloat handles the string elapsed field correctly

🤖 Generated with Claude Code