[aliromman]

The invoice comparator accepts declared histograms, so a missing target debit or a transcription error can distort a proposed kappa improvement. This adds an opt-in exporter for the selected P104 source. It authenticates the selected closure, reconstructs all local, query, idle and main charges, and requires agreement with the reviewed component and complete receipts before it emits the common invoice schema.

The result is 66,726 paid calls, 13,048 productive fields and rank surplus 760. The evidence output binds the exact invoice bytes and each source input. The current selected result and frozen sources are unchanged. This addresses the exporter request in owned-relation-exhaustion.

python3 -B scripts/export_selected_invoice.py --evidence /tmp/invoice-evidence.json > /tmp/invoice.json
python3 -B scripts/compare_recursive_invoices.py /tmp/invoice.json CANDIDATE.json
python3 -B -m unittest discover -s tests -p 'test_research_selected_invoice.py' -v

Verification: seven focused tests pass on Python3.14.7 and registered Python3.12.14, also with -I -O -B. They use an independent two-path formula, remove each physical charge class and each paid/target rank entry, test ownership and opcode failures, corrupt each actual source file, remove an input from the authenticated inventory, and exercise the CLI-to-comparator path. All ten existing comparator tests, selected baseline binding and git diff --check pass. New tests join the existing test_research_*.py CI discovery.

Exporter, tests, usage and scope.

Source: selected audited head 09dd0965711ca7eeb169057a420abab1c9052137, selected conditional kappa0.00310045. The P104 word is by Pierre-Loic Doulcet (@hexapode); Ali Romman supplied the complete caller and helper accounting; Cinna Davis supplied the controller. Jacob Alan Sussman supplied productive recursion. The comparator and common schema are by zxv. New extraction and tests: Ali Romman with substantial OpenAI Codex assistance, Apache-2.0.

Limits: accounting reconstruction and source binding do not replay scalar endpoint matrices, native phases or compiler inequalities. Candidate invoices still need their own complete justification. Unknown source formats fail explicitly. The full selected theorem replay and independent review were not run. This shallow checkout does not support full-history preflight. Hosted CI is pending; main's two known selected-navigation failures are already covered by PRs #5659 and #5703.

Coordination: canonical claim. This is a documented handoff, with an explicit change from the sampled replay-runtime card. Main, priority7 and selected-result bindings were refreshed before work and submission. No frontier update or merge is requested.