Prime Agent handoff follow-up acceptance

The operator authorized three fixes after the Prime Agent session report: kb/messages/20260905T191941Z-codex-orchestration-prime-agent-session-report.md. The citation clarification is explicit: quoted source code must be verified by full-text search, not by trusting manually assigned line numbers.

Implemented contract

1. The run-state type includes a copyable Git source mapping with path and sha256: null. Startup requires validation after filling the source, before inspection or delegation. This does not add an initializer or phase ledger.
2. Main and specialist procedures require minimal verbatim source excerpts for load-bearing findings. Code supports implementation claims; source prose can support claim-level findings. The parent retains each passage once on its canonical record. Existing publication checks search the whole pinned blob or immutable capture after whitespace normalization. Display line numbers, invented ellipses and formatting fences are excluded from quoted text. Complete results and memory reports with no quote anchors now fail standing validation. This catches the prior total omission of quoted evidence; it does not establish coverage or semantic support for every material claim.
3. Handoffs spell every identifier in full. Integration maps exact tokens after expansion, without substring replacement or implicit ranges. Standing result validation checks declarations and canonical references throughout ordinary prose, not only comparison fields. It rejects duplicate record declarations, unresolved canonical references, unintegrated proposal IDs outside Reconciliation, and shorthand/ranges. Source quotations and fenced evidence are excluded from identifier scanning. Memory reports reject shorthand; their comparison references retain the existing commissioned and proposed-record checks.

These requirements apply to the current workflow without an old-format fallback. Completed generated analyses are not patched to pass the new rules. The Prime Agent output remains the session's original evidence, not a newly attested result under this contract.

Bounded verification

Regression cases include the reported OBJ-15/O2/O3/O4/O5 and RTE-20–R9, a shortened tail after two full IDs, complete lists, unresolved references outside comparison fields, duplicate declarations and unrelated IDs inside source quotations/fences. An integration test proves the standing result validator rejects corrupted prose references. The corrected Prime Agent result passes the new identifier scan; that check does not retroactively certify its source evidence under the restored quote requirement.

A publication fixture without quotes now fails for either its main result or memory report. A pinned source-code fixture proves that quoted code is found outside the attributed navigation line, while fabricated text present only in the changed worktree is rejected. The Git source example passes actual running state validation after substitution of the fixture repository and revision.

All 769 tests pass. Repository Ruff, changed-document validation and diff checks pass. This is a machinery and contract check, not a fresh production analysis or semantic audit of Prime Agent.