v0.18.0

Batched release of everything since v0.4.1 — the loop's autonomy + the self-improving flywheel.

Loop reliability

• Same-tier CI-fix budget before escalating (v0.12.0) — a red check gets ci_fix_max same-tier fixes (error injected) before climbing a model tier, so a one-line lint nit no longer burns smart→reasoning→opus.
• Pre-PR local gate (v0.13.0) — runs the repo's real checks in the worktree and hands the coder the failure to fix before opening a PR; opens anyway after the budget (CI stays the backstop).
• Auto-rebase on conflict (v0.14.0) — a sibling merge that stales a PR is rebased onto base in a detached worktree (clean → force-push, no coder; conflict → bounded coder re-resolve).
• Specific CI failure recorded (v0.16.0) — the attempt comment carries the real error (golden-map / F841 / schema), not the Failing checks: header.

Self-improving flywheel (the loop learns from itself)

• loop-retro (v0.15.x) mines the board's attempt/outcome history into recurring failure classes + flow stats (board_retro tool + retro.py).
• Distill → grounding (v0.17.x) — schedule-safe, propose-over-create: writes a dated retro report and (v0.18.0) writes lessons to the knowledge graph (loop-lessons).
• Inject into coders (v0.18.0) — the loop reads relevant loop-lessons from the KG (graph.sdk.knowledge_search) and injects a "Known gotchas for this area" block into each coder's prompt, so known failures are heeded on attempt 1.

Plus earlier autonomy/resilience work (goal-verify gate, keep-worktree retries, format fixups). Releases now run on the fleet protoLabsAI/release-tools cadence (LLM notes + Discord).