Share request evidence and replan missing analysis prerequisites

[mikamikasuki]

Changes

• Return MISSING_PREREQUISITES before select/compare opens a missing or unsettled source receipt. The standalone read-only classifier reports existing work IDs, config references, supervisor state and actionable next operations. The Agent chooses the scientific operation.
• Reuse the existing episode/native replanning feedback path. A failed analysis can report newly checked missing sources while its original failed row stays intact; successful cached actions and failed GPU work retain their existing semantics.
• Intern repeated JSON across the complete request in one reversible shared reference pool. Preserve original message roles, boundaries, metadata, literal text, JSON spelling, evidence identities and response schemas. Complete scientific originals remain in persistent storage and accessible through existing context reads.
• Replay an unchanged completed native request across the presentation upgrade without another provider call. Reject source/settings changes and preserve unsettled-request reconciliation.
• Document the definition/reference and persistent-artifact patterns used by Pydantic and LangChain; no new dependency or execution framework.

Validation

189 focused tests passed in 11.66 seconds:

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

The saved failing request is tested read-only with exact message restoration and unchanged schema/source records. Its actual CLI wire shrinks from 1,096,372 to 749,333 characters, below the observed 1,048,576-character interface capacity. A real signed-in Codex CLI call with gpt-6-luna / medium completed in 5.91 seconds and passed the original decision Schema. This transport diagnostic executed no scientific operation and is separate from formal results.

DEV orchestration tests explicitly use nonformal fixtures. They verify missing feedback, fresh Agent decisions, six source identities, valid-receipt selection/comparison, native reload, no phantom pending action, and no silent rerun of an already-failed selector. Formal inference is validated during existing-episode recovery using the original supervisor, ledger, configs and real artifacts.

Deployment recovery

Existing campaign records, failed requests, decisions, usage and scientific artifacts remain intact. Restore the two affected parent work IDs; retain the failed selector while missing-source feedback reaches the Agent. Only after all six genuine inference executions and their configuration-bound receipts/artifacts are reconciled can the existing failed CPU selector be explicitly restored under its original work ID, preserving its prior failure record.

No cumulative Agent Token/call/step/time limit, new blocking state, automatic scientific prerequisite execution, new ledger, experimental data or scientific-protocol change is included.