JVNAUTOSCI-2600: harden adaptive provider and effect seams

[witbrock]

Status

This is a stacked, draft successor to #311. It is not an accepted candidate: the disclosed regression gate was partial/pass/fail, so no fresh holdout, shadow work, A4, merge, cutover, or release was attempted.

What changed

• Resolve answer-only and structured OpenAI calls against the exact represented API profile, including profile-scoped temperature and reasoning constraints.
• Preserve conservative Chat fallback without borrowing another connection's parameter policy.
• Atomically admit effects only when their full configured hard window remains; expose bounded, actor-scoped late-effect receipts and freshness.
• Harden late-effect idempotency and actor-scope collision handling.
• Forward the complete adaptive transport interface through the restricted von_chat_run gateway.
• Batch concept text enrichment while preserving completeness when text relations are dangling or capped.
• Reject explicitly individual concepts as semantic parents for non-predicate create_concepts writes.

Why

The A3r4 path exposed three generic support-surface defects: required Responses profiles were lost on answer-only calls, timed-out effects could complete after the turn without durable actor-visible evidence, and concept enrichment paid avoidable Atlas read costs. Disclosed A3r5 replay then exposed a fourth exact-interface defect: an organisation individual could be used as a type parent, producing malformed durable scenario concepts.

These changes remain mechanism and persistence work. There is no ecology-, paper-, persona-, model-name-, or case-ID-specific semantic policy in Python.

User/developer impact

Provider calls fail or sanitise honestly against represented authority, late effect outcomes can be diagnosed through canonical read-back, clipped writes are not silently dispatched, and malformed individual-as-type parenting is rejected before mutation.

Known blockers remain: cold model-registry hydration can exhaust the first caller deadline; generic 20-second effect admission starves later writes in sequential batches; late-receipt submission is an in-process best-effort queue rather than a durable outbox; and live create_concepts/concept-resolution Atlas latency remains high.

Validation

• Provider transport/temperature: 72 passed.
• Adaptive effect/deadline/restricted-gateway/turn-record: 128 passed.
• Model parameter/registry: 25 passed.
• Create-parent/duplicate/batching/blocked-parent: 29 passed.
• Focused final repairs: 12 passed.
• Ruff, Python compilation, and git diff --check: passed.

The broader turn-execution read-tool file had 28 passes and 11 existing live-authority expectation failures (workflow_global_admin_authority_required); a representative failure reproduces on the frozen A3r4 base, while the new late-observation tests pass. No full-suite or acceptance claim is made.