Throttle concurrent pipeline AI steps

[chubes4]

Summary

• Adds a site/provider-aware lease limiter for pipeline AI execution so concurrent fetch lanes do not stampede provider transports.
• Defers above-limit AI steps by rescheduling the same datamachine_execute_step action and parking the job back to pending without marking work failed or processed.
• Adds focused smoke coverage for acquire/throttle/release behavior, non-AI step isolation, throttle metadata, and retry-policy preservation.

Tests

• vendor/bin/phpcs inc/Core/Steps/AI/AIStep.php inc/Engine/AI/PipelineAIConcurrencyLimiter.php inc/Engine/AI/PipelineAIConcurrencyLease.php inc/Abilities/Engine/ExecuteStepAbility.php inc/Core/PluginSettings.php tests/ai-step-backpressure-smoke.php
• php -l inc/Engine/AI/PipelineAIConcurrencyLimiter.php && php -l inc/Engine/AI/PipelineAIConcurrencyLease.php && php -l inc/Core/Steps/AI/AIStep.php && php -l inc/Abilities/Engine/ExecuteStepAbility.php && php -l tests/ai-step-backpressure-smoke.php
• php tests/ai-step-backpressure-smoke.php
• php tests/job-retry-policy-smoke.php

Caveats

• Verified with focused pure-PHP smoke coverage; not deployed and no live 3-flow burn-in was run from this worktree.

AI assistance

• AI assistance: Yes
• Tool(s): OpenCode (openai/gpt-5.5)
• Used for: Adding pipeline AI concurrency/backpressure and tests. Chris remains responsible for review and merge.