v3.10.0 — reusable strategy quality becomes an injection gate

Reasoning-pattern confidence measures cluster support, not whether a strategy is safe and useful to transfer into an unrelated task. Version 3.10.0 separates those two concerns.

Highlights

• LLM naming now returns reusable, quality_score, and factual quality_reasons.
• The naming prompt removes names, clients, projects, repositories, paths, filenames, URLs, identifiers, run numbers, and incident-specific values while requiring complete, self-contained steps.
• Deterministic validation catches generic mechanical titles, raw Moves: output, unfinished prose, and recognizable context-specific identifiers.
• reasoning_training.injection_min_quality adds an independent injection threshold; the default is 0.7.
• A reversible per-pattern injection override is available through the dashboard and PATCH /api/dashboard/reasoning/patterns/{pattern_id}/injection.
• Dashboard, CLI, MCP, and REST pattern listings expose quality, reusability, naming method, reasons, and injection state.

Compatibility and rollout

• Existing patterns without quality metadata retain their previous eligibility until they are rebuilt or manually disabled.
• New heuristic-fallback patterns remain inspectable but are not injected automatically.
• No pattern or reasoning trace is deleted by this upgrade.
• Publishing this draft should happen only after PR approval and merge. Production deployment remains a separate, backed-up atomic rollout.

Verification

• 7,367 tests passed, 132 skipped, and 1 expected failure in the clean non-stress suite.
• Ruff, formatting, mypy, dashboard build, scoped ESLint, Vitest, and reasoning Playwright E2E passed.
• All 16 version-bearing fields are synchronized at 3.10.0.