Feedback session cost driver is transcript length not startup

An audit of Justin's actual session logs (4,117 sessions) found that startup overhead (system prompt, tool schemas, CLAUDE.md, MEMORY.md, MCP schemas) is only ~7% of total cost (~$437), and CLAUDE.md+MEMORY.md together are only ~4% of a typical per-turn cache read (~7,860 of 184k tokens). The real driver: 6% of sessions (250, each 31+ turns) account for 91% of all spend, because cost scales roughly quadratically with turn count — a session dragging its full history through cache on every turn. One 1,450-turn session cost $619 vs. an estimated $116 if compacted into chunks; late turns in it cost 7.1x more than early turns.

Why: Justin asked for a real audit rather than guesses, after initial advice (trim CLAUDE.md, disconnect MCP servers) was measured and found to be chasing the wrong 7%.

How to apply: When advising Justin (or anyone) on reducing Claude Code token/cost usage, do NOT recommend trimming CLAUDE.md/MEMORY.md or disconnecting MCP servers as the primary lever — that's a rounding error. Instead recommend /compact on long-running working sessions periodically, /clear at hard topic boundaries, and breaking marathon work into multiple focused sessions instead of one very long one. Repo/archive size on disk is irrelevant to token cost since files load only on demand.

[auto-memory session 868e85f6-8948-4d2c-a415-6388ca0ec5a3, confidence 0.85, mode direct]