Deliver actionable analysis and context feedback after restart

[mikamikasuki]

Changes

Make existing Runtime feedback actionable for the next Agent decision after restart. This PR is stacked on #296; it changes existing validation/analysis/context call sites without adding a state machine, dependency framework, execution limit or automatic scientific work.

• For an exact known work ID with conflicting optional selectors, report supplied and expected fields and the existing copy-or-null/omit rule. Conflicting proposals remain rejected.
• For the existing source-selection comparison guards, report the observed selector state/result and matching registered next-action selectors. The first guarded action still stops the batch; later proposed actions remain unexecuted.
• Propagate current child feedback through the existing aggregate response_validation_feedback field. Preserve valid prefix results and earlier actual decisions.
• Distinguish invalid public-memory queries from source IO, permission, capacity and measurement failures. Invalid queries request a new Agent decision and provide the exact query/error, authorized segment IDs and root-pointer syntax; they do not create a successful page receipt or scientific progress.
• Refresh the generated public catalog to match current bindings without changing previous messages or persistent source versions.
• Use native AIDE's structured exc_info.args shape so a validation diagnostic survives Journal restart.

The prerequisite classifier remains in #296. Both patches reuse existing supervisor states, registered work identities, execution receipts and the original external ledger. No candidate selection, inference, comparison or corrected query is automatically chosen for the Agent.

Validation

232 focused tests passed in 15.74 seconds, with the original saved incidents selected read-only and no skips:

tests/test_context_query_feedback.py
tests/test_source_selection_feedback.py
tests/test_action_binding_feedback.py
tests/test_study_analysis_prerequisites.py
tests/test_shared_message_context.py
tests/test_aide_runtime.py
tests/test_study_runtime.py
tests/test_study_context.py
tests/test_context_delivery_completion.py
tests/test_codex_transport_diagnostics.py

Coverage includes complete-batch validation, current guard feedback after ordinary Runtime and native Journal restart, real deterministic CPU comparisons from explicit nonformal fixtures, valid-prefix/blocked-suffix ordering, invalid pointer recovery, exact original-source reads, no false delivery/coverage, original-request preservation, provider-boundary measurements and unchanged pending GPU reconciliation.

Four actual selector-conflict responses and the original absent-pointer response are revalidated without modifying their requests, responses, source segments or usage records. Nonformal fixtures are explicitly ineligible for scientific scoring and do not perform model or GPU inference.

Campaign recovery

Affected episodes resume through existing reconciliation/resume paths on the same ledger. Original failures, proposals and resource records remain preserved. The six actual inference jobs and their successful receipts remain unique; their raw predictions were independently checked. A real signed-in Codex Agent subsequently chooses any eligible comparison or corrected read.