Add session token efficiency scanner

[cbusillo]

Summary

• add a deterministic read-only scanner for Every Code rollout/session token-efficiency suspects
• report peak cumulative tokens, largest turns, cache ratio, token resets, image payloads, large persisted records, duplicated instruction blocks, and optional usage correlation
• document the no-LLM-by-default policy and add a focused synthetic rollout test

Evidence

Running against recent May 11 sessions identified:

• 152.7 MB of persisted data:image payloads in a 165 MB rollout
• duplicated project-doc/skill injection blocks across multiple sessions
• several 200k+ token single-turn events

Validation

• python3 -m unittest tools/session-token-scan/test_scan.py
• python3 -m py_compile tools/session-token-scan/scan.py tools/session-token-scan/test_scan.py
• python3 tools/session-token-scan/scan.py ~/.code/sessions/2026/05/11 --limit 20 --top 5
• python3 tools/session-token-scan/scan.py ~/.code/sessions/2026/05/11/rollout-2026-05-11T18-37-18-3f238f4d-f245-4f65-aa53-4dee95ba7758.jsonl --json --top 2
• ./build-fast.sh

Closes #44