[witbrock]

Outcome

This is the single main-targeting candidate for the thin adaptive Von. It combines the complete A3r8 revision stack with current main, the controlled-evaluation boundary record, and durable dismissal of failed prompt-queue rows.

The previously merged JVNAUTOSCI-2605/2606 UI fixes are already in main. The equivalent stacked A3 UI commit has the same patch-id, and the clean aggregate merge preserved the current-main implementation without duplicate resolution drift.

Included

• remove the universal ordinary-turn controller and use the thin adaptive turn path;
• retain bounded read/effect seams, durable receipts, identity/scoping, and honest incomplete-effect reporting developed through A3r8;
• retain current-main attachment-state and single-renderer table fixes;
• allow failed queue records to be dismissed durably while preserving failure evidence and never rerunning them;
• record the controlled-evaluation boundary without claiming unavailable C0/D0 fixtures.

The stacked draft PRs #309-#317 remain open temporarily as revision history. This aggregate PR is the candidate to review against main.

Validation

• 281 focused backend tests passed across adaptive turn, queue, identity, relationship, display-element, and execution-summary surfaces;
• 105 additional transport, deadline, model-registry, evidence-store, and relation-index tests passed;
• 30 frontend Jest tests passed across queue actions, attachment workflow input, display-element ownership, and Markdown/table rendering;
• frontend static lint, JavaScript syntax, Ruff, bytecode compilation, and git diff --check passed;
• clean AgentTest runtime on port 5016 reported /health for exact commit b4b501164949de75c7a7ebdd877d7f3d5bc5020f with model-registry authority ready;
• signed-in browser replay completed a real read-only recovery turn with gpt-5.6-luna in 101.5 seconds, used only fetch_concept, find_relations_with_argument, turn_capabilities, and turn_read_evidence, reported canonical state without represented-state mutations, rendered one seven-row table exactly once, and emitted no browser console warnings/errors;
• GitHub minimum Python syntax, secret scanning, JavaScript/TypeScript CodeQL, and Python CodeQL all passed.

Known limits

• one broad find_relations_with_argument call hit its 20-second deadline before the model recovered with narrower parallel reads; the completed turn still took 101.5 seconds and exposed Atlas query work worth narrowing;
• the controlled C0/D0 fixtures are still unavailable, so this PR does not claim completion of the stronger comparative research evaluation;
• this is a large aggregate change (240 files, approximately +39.6k/-109k), reflecting deliberate subtraction of the former controller and its expectation tower. Review should focus on the end-to-end candidate and regression evidence rather than preserving removed architecture by default.

The PR is ready for review. No merge or production cutover has been performed.

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.