chore(release): agent-runtime 0.45.0

[drewstone]

Cuts the 58-commit backlog on main (v0.44.0..HEAD) into a published release. This single release unblocks three things at once:

1. agent-app's loop un-fork — runToolLoop/streamToolLoop (feat: bounded turn-level tool-dispatch loop (runToolLoop / streamToolLoop) #137) ship, so agent-app's hand-rolled runAppToolLoop collapses to a 1:1 re-export.
2. The RSI proof eval — the recursive Agent.act/Supervisor machinery + improvementDriver (the optimization API collapse onto agent-eval selfImprove, refactor(improvement): collapse optimization API onto agent-eval selfImprove #172) become importable so compareDrivers can race baseline-vs-RSI.
3. Agentic headroom corpus — AppWorld/commit0/aec-bench/EnterpriseOps-Gym deployable adapters (feat(bench): wire real benchmark adapters — aec-bench, commit0, programbench, appworld + shared harness #153/feat(bench): benchmark unifier — runBenchmarks over one ADAPTERS registry, AgentProfile the only knob #156/feat(bench): EnterpriseOps-Gym adapter — deployable SQL state-checker (the gate's non-coding middle-band domain) #157) ship, where lift is actually measurable.

Headline surface

• runToolLoop / streamToolLoop — bounded turn-level tool-dispatch loop (feat: bounded turn-level tool-dispatch loop (runToolLoop / streamToolLoop) #137)
• RSI agent tree: recursive Agent.act, Supervisor keystone, runProgram, adaptive-driver channel (feat(loops): runtime steer-firewall + dynamicLoopRunner analyze forwarding (RSI Gen-1) #139/feat(loops): recursive execution atom — budget-conserving Scope + Supervisor keystone #151/feat(runtime): gen-6 architecture consolidation — one recursive agent tree, observable + deep-cleaned #165)
• optimization API collapsed onto agent-eval selfImprove; runtime keeps the CODE-surface ImprovementDriver (refactor(improvement): collapse optimization API onto agent-eval selfImprove #172)
• benchmark adapters + runBenchmarks over one ADAPTERS registry (feat(bench): wire real benchmark adapters — aec-bench, commit0, programbench, appworld + shared harness #153/feat(bench): benchmark unifier — runBenchmarks over one ADAPTERS registry, AgentProfile the only knob #156/feat(bench): EnterpriseOps-Gym adapter — deployable SQL state-checker (the gate's non-coding middle-band domain) #157)
• agent-eval floor raised to >=0.83.0 (docs: sync optimization docs to selfImprove + bump agent-eval floor to 0.83 #175)

Verification (full publish.yml verify gate, run locally on this tree)

INSTALL(frozen) ✓ LINT ✓ TYPECHECK ✓ TEST ✓ BUILD ✓ VERIFY:PACKAGE ✓

Tag v0.45.0 will be pushed after merge to trigger the OIDC publish (version-lock: tag == package.json).