Enumerate 1890 warehouse ledger reconstructions from OCR candidates

Reconstruct a small fictional 1890 warehouse ledger without silently resolving ambiguity. Deliver normalized.csv, audit.md and a standard-library verifier, not a web application. The unit is boxes. Reliable opening stock is 40. Five consecutive signed movements t1..t5 have OCR candidate sets: t1 receipt {+12,+17}; t2 dispatch {-9,-4}; t3 receipt {+6,+8}; t4 dispatch {-15,-13}; t5 receipt {+10,+16}. Reliable checkpoint: stock after t3 is 54; reliable final stock after t5 is 51. Enumerate ALL candidate reconstructions satisfying both checkpoints and nonnegative intermediate stock. Report which movements and balances are identifiable and which remain ambiguous. normalized.csv must have one row per t1..t5 with columns id,kind,candidate_deltas,identified_delta,possible_balance_after; leave identified_delta empty when multiple values remain, and do not pick a preferred reconstruction. audit.md must preserve raw candidate sets separately from surviving alternatives and explicitly distinguish a resolved value from a plausible guess. Include all surviving five-movement sequences in the report, with arithmetic. The verifier must derive survivors from raw candidate sets and checkpoints, compare them against the delivered CSV/report claims, and exit nonzero on disagreement. No invented source documents or historical assertions. Return the key reconstruction and ambiguity conclusions directly in the final response as well.

Repair the existing historical ledger deliverable, chiefly verify_ledger.py. Preserve the supplied raw candidate data and the correct two surviving reconstructions. An independent audit found two checker defects: changing audit.md from 'Exactly two sequences survive:' to 'Exactly three sequences survive:' still exits 0, while replacing the correct phrase listing '+12, +17' with '+12 and +17' exits 1. Fix both semantically within an explicitly documented, bounded report format; do not claim arbitrary natural-language verification. Add repeatable regression checks that operate on isolated copies: original passes; that false count fails; equivalent ambiguity wording passes; normalized.csv t3 identified_delta +6 changed to +8 fails; falsely identifying t1 as only +12 fails. Keep audit.md accurate and reader-facing and preserve unrelated correct content. Run and report the checks and their exit codes. Deliver the updated offline files, not a website.

Repair the residual verifier defect in the just-delivered ledger revision. Independent testing changed ONLY the actual claim under '## Ambiguous movements and balances' from 't1 remains ambiguous between +12, +17' to '+12, +99', leaving the grammar documentation unchanged: verify_ledger.py wrongly exits 0 because it searches the whole document. regression_checks.py also replaces the first occurrence, which is now a documentation example, not the real field. Scope every report check to its designated unique section/field, parse values and compare with derived survivors, and reject missing/duplicate/conflicting actual fields. Documentation examples must neither prove nor invalidate actual claims. Preserve correct raw data, CSV and both survivors. Keep a bounded documented format rather than claiming natural-language understanding. Extend isolated-copy regressions: original 0; actual t1 comma to and 0; actual t1 +17 to +99 nonzero while examples remain; actual ambiguity claim removed but examples retained nonzero; duplicate conflicting actual t1 claim nonzero; survivor count two to three OR four nonzero; wrong t3 CSV and false identified t1 still nonzero. Each mutation must assert that exactly its intended real field changed, not an example. Run all tests after final document changes. Deliver updated offline checker, harness and accurate audit.