[backtrack5r3 (tomflenner)]

What does this PR change?

KTN-201 generated write-only AI union and response models, reserve/commit/run helpers, typed errors/keyed retries, model and conversion-rule management, inference/reservation reads, optional OpenAI/Anthropic drop-in/delegation wrappers with sync/async streaming settlement. Adds extras, CI compatibility checks, changelog and version 1.1.0.

Testing

Integrated full suite: 948 passed, 8 optional provider compatibility tests skipped because provider packages absent in main environment. Isolated provider-compatible focused suite previously passed with real installed SDK types and synthetic transports. Ruff, mypy (76 files), model/sync generation freshness and targeted contract tests passed.

Compatibility / migration impact

Vendored contract from companion Kaiten feature branch; source.yaml records source commit. Strict read union remains non-AI. AI handwritten response classes replaced with compatibility aliases. Providers are optional extras, no prompts sent to Kaiten.

Decisions to review

NOT_SENT requires demonstrable pre-send failures; ambiguous post-send errors remain provider errors. Usage extraction/validation failures preserve provider results and attempt settlement without usage. Agent interfaces do not establish awaited veto-capable preflight before every internal model inference; full agent wrappers were not fabricated. This is an explicit scope gap pending API/requirements decision.

Not in this PR

Provider proxy/server pricing, universal agent preflight, tags/publication.

Checklist

• Tests/docs/extras/attribution updated; commits signed off
• No provider keys, prompts or customer data
• Optional compatibility tests on final main environment / CI green

[backtrack5r3 (tomflenner)]

Naming refactor: high-level AI helpers/meters and OpenAI/Anthropic wrappers now use entitlement= instead of pool=. Low-level entitlement_slug, HTTP routes and connection-pool settings are unchanged; old high-level pool= is rejected. Updated README/changelog/examples and regenerated sync code. Local verification: 988 tests passed, 8 optional provider compatibility tests skipped; Ruff, strict mypy and generation freshness passed.

[backtrack5r3 (tomflenner)]

Dedicated AI change and billing dependency

This PR remains dedicated to the AI credit control plane; AI fixes have been published here, not on the billing branch. It depends on billing PR #5 and should land after billing.

The AI and billing branches diverge in generated contracts and shared transaction-error definitions. A base-only retarget would be misleading. Integrate billing's merged base before merging this PR, retaining one shared error class and preserving numeric report versus AI-phase replay semantics. No force-push was performed.

Latest verification: 541 focused tests and Ruff passed, including inherited-header suppression and exact usage validation. Earlier broader verification had optional-provider skips and one missing-OpenFeature failure; final full-suite/mypy and typed contract refresh remain incomplete.