fix(optimize): exclude sidechains from behavioral signals

[avs-io]

Problem

Claude subagent sidechain transcripts were treated as user-started sessions by codeburn optimize. They inflated the displayed session population and generated low-worth, context-heavy, outlier, capability, coaching, and model-default advice that sidechains fail by construction.

Root cause

The session-level isSidechain marker was parsed but not consistently applied across optimize consumers. Some raw transcript detectors classified each line independently, the compact parser for large JSONL lines dropped the marker, and model/coaching paths consumed sidechain-inclusive project populations.

Change

• Use one user-started-session predicate for optimize session counts and behavioral projections.
• Exclude sidechains from per-session findings, raw read/edit detectors, capability reliability, coaching, churn, and model-default recommendations.
• Keep original projects for aggregate spend, calls, tokens, cost-rate, and configuration-overhead findings.
• Make raw Claude transcript classification sticky across the whole sidechain file, including earlier unmarked calls.
• Preserve isSidechain in the compact parser for JSONL entries larger than 32 KiB and through warm-cache/range rebuilds.
• Include sidechain classification in optimize cache identity.

User impact

codeburn optimize reports user-started session counts and no longer recommends behavioral or actionable changes based solely on subagent transcripts. Delegated token usage remains fully represented in spend, call, token, and configuration-overhead totals.

Preservation and out of scope

• Sidechain spend is not deleted or attributed to the parent.
• Overview, general model-efficiency, and global compare populations are unchanged.
• Unknown or explicitly false markers remain user-started under the existing contract.

Testing

• Credible composition RED: sidechain-only projects previously emitted an actionable model downgrade and correction coaching despite reporting zero optimize sessions.
• Large/sticky classification RED: 3 expected failures proved large lines lost the marker and earlier calls escaped late transcript classification.
• Focused matrix: 116/116.
• Full root suite: 2,632 passed, 5 skipped.
• Serial lock suite: 26/26.
• Full desktop suite: 468/468.
• Root, dashboard, and desktop TypeScript checks passed.
• CLI, dashboard, and desktop production builds passed.
• Exact-SHA independent spec and conservation reviews approved.
• Isolated real Claude discovery QA: 0 sessions, 67 calls, $0.210 retained, no model recommendation, no coaching, and no sidechain-only behavioral findings.
• Real TUI and rebuilt Electron checks preserved aggregate spend while Optimize showed only an unrelated configuration finding.

Fixes #974.