Upgrade to RubyLLM 1.16 and adopt new capabilities

[adham90]

Summary

Bumps the dependency to ruby_llm >= 1.16.0 and adopts four capabilities that landed in RubyLLM 1.15/1.16, closing real gaps in cost analytics, tooling, and instrumentation. Fully backwards compatible — no schema migration required.

Feature	What it does
Accurate cost	total_cost now prices prompt-cache reads, cache writes, and reasoning/thinking tokens via RubyLLM::Cost on top of authoritative (attempt-aggregated) text tokens. Per-component breakdown stored in executions.metadata["cost_breakdown"]. The async ExecutionLoggerJob no longer recomputes a text-only total over an accurate one.
Concurrent tools	New inheritable tool_concurrency DSL on BaseAgent (false/true/:threads/:fibers); passed as concurrency: to with_tools when set, else RubyLLM's global default.
HTTP instrumentation	Subscribes to RubyLLM's request.ruby_llm events during the LLM call and records real provider latency + count into executions.metadata (llm_request_ms, llm_request_count), distinct from total pipeline duration.
Config passthroughs	Forwards bedrock/mistral/perplexity/vertexai/xai_api_base, faraday_adapter, deprecation_behavior, and tool_concurrency to RubyLLM.config.

Already-supported 1.15/1.16 features (image editing, transcription word timing, additive callbacks) needed no changes.

Design notes

• No schema change. The primary instrumentation path persists context.total_cost directly, so making it accurate flows straight through the existing total_cost column. The cache/reasoning breakdown lives in metadata rather than new columns, avoiding a host-app migration. Promoting the breakdown to dedicated cost columns + dashboard display is noted as a follow-up.
• Text input/output cost is unchanged (priced from context tokens); cache/reasoning is additive and degrades gracefully when the registry lacks a price.
• Also deletes an orphan WIP spec (attachments_spec) that referenced a nonexistent DSL::Attachments module and crashed the suite at load time.

Testing

• Full suite: 4689 examples, 0 failures (+19 new specs across the four features plus an end-to-end integration test that runs a real agent through the pipeline and asserts the cache-aware total persists to the DB).
• standardrb clean.
• Line coverage 90.65%.

Docs updated: CHANGELOG, README requirement, and the Configuration / Agent-DSL / Tools / Pricing / ActiveSupport-Notifications / Execution-Tracking wiki pages. An internal ADR is in changelog/.

🤖 Generated with Claude Code