[Epic] v4.2 Roadmap — harden + make smarter

[egorfedorov]

v4.2 Roadmap — "harden the honest optimizer, then make it smarter"

Tracking epic for the next milestone. CCO v4.1 made savings honest and quiet; v4.2 should make the underlying mechanics trustworthy under real conditions (parallel hooks, Windows, MCP-heavy sessions) and then smarter (shared learning, self-tuning). Grouped by priority.

🛡 A — Correctness & data integrity

• Data race: concurrent hooks lose updates on shared session ledgers #32 — Data race: concurrent hooks lose updates on shared session ledgers
• Windows support: CRLF in .contextignore + path-separator assumptions #33 — Windows support: CRLF in .contextignore + path-separator assumptions

🎯 B — Accuracy of the core metric (token estimation)

• Per-language token estimation (replace global AVG_CHARS_PER_LINE=35) #35 — Per-language token estimation (replace global AVG_CHARS_PER_LINE=35)
• Budget: use actual file size for Read estimate instead of 2000-line default #34 — Budget: use actual file size for Read estimate instead of 2000-line default

✅ C — Test coverage of decision logic

• Test the untested decision logic: tracker finalization, read-cache staleness, budget gating #36 — Test tracker finalization, read-cache staleness, budget gating

✨ D — New capabilities

• Feature: cross-project / team pattern sharing (export & import patterns digest) #37 — Cross-project / team pattern sharing (export & import patterns digest)
• Feature: self-learning MCP/Agent token estimates from observed result sizes #38 — Self-learning MCP/Agent token estimates from observed result sizes
• Config surface + /cco-config: expose hardcoded thresholds #39 — Config surface + /cco-config: expose hardcoded thresholds

📦 E — Distribution

• Docs: add real CHANGELOG.md and CONTRIBUTING.md #40 — Add real CHANGELOG.md and CONTRIBUTING.md

---

Suggested sequencing

1. A first (Data race: concurrent hooks lose updates on shared session ledgers #32, Windows support: CRLF in .contextignore + path-separator assumptions #33) — silent correctness bugs erode trust fastest.
2. B + C together (Per-language token estimation (replace global AVG_CHARS_PER_LINE=35) #35, Budget: use actual file size for Read estimate instead of 2000-line default #34, Test the untested decision logic: tracker finalization, read-cache staleness, budget gating #36) — fixing the metric and testing it go hand in hand; this is the "honest savings" core.
3. D (Feature: cross-project / team pattern sharing (export & import patterns digest) #37–Config surface + /cco-config: expose hardcoded thresholds #39) — build new features on a trustworthy base.
4. E (Docs: add real CHANGELOG.md and CONTRIBUTING.md #40) — ship with a real changelog.

Each item is independently shippable; no item blocks another except that C (#36) is easier to land alongside B.