[kumagi]

Summary

PR2977 checker parity vs exporter membership
  + exact agreement characterisation (all positive multiplicities odd)
  + guard comparison and selected-producer row-space check
= the schema decision is uniqueness-as-producer-restriction, not semantic necessity

Adds only research/scatter-parity-contract-audit/. It independently answers the boundary question this card names: for the frozen xcheck_p.py scatter loop (one XOR per occurrence) and export_records.py scatter emission (one ADD per membership), the two operators agree on a word exactly when every positive multiplicity in every scatter row is odd. The proposed simple-support guard is sound — it never admits a divergent row — but strictly stronger: on all 1,365 tuples of length 0..5 over four retained-total indices it rejects 296 rows that agree, the smallest being [0,0,0]. The PR2977 mutant [4,5,6,7,8,0,0] diverges and both contracts reject it. The pinned tuned word has 111 simple, ascending rows (527 terms = 527 membership ADDs), so either contract preserves the selected bytes.

Task supplier-boundary-audit, priority revision 6, base 836b19bfa309ceafbc8ee32a15e8c4b10dde7e11 (current main; no rebase needed). Canonical thread: discussion #2976.

Evidence

• Source-line binding: checker for k in row: val[v + t] ^= copies[k], exporter if k in row: rec.append(...), generator decoder_rows from itertools.combinations, tuner rewrites only A/C (the string scatter does not occur in it). The tuned word is bound by SHA-256 e3e301e6….
• Exhaustive domain (1,365 rows): divergent 1,004; simple 65; odd-multiplicity admissible 361; agreeing rows the guard rejects 296; occurrence-count vs emitted-ADD mismatches 1,300. Controls: membership-loop substitution makes the mutant agree (isolates the divergence to the occurrence loop); simple ⊆ exact and divergent ∩ simple = ∅ hold on the whole domain.
• Accounting observation: scatter_terms counts occurrences while the exporter emits membership; the counts coincide on simple supports (527 = 527) and differ on every repeated-index row, including the 296 agreeing ones — a schema must say which is authoritative if odd multiplicities are ever admitted.

python3.12 scripts/research_package.py --package research/scatter-parity-contract-audit
# PASS_SCATTER_PARITY_CONTRACT domain=1365 divergent=1004 simple_accept=65
#   exact_accept=361 over_rejected=296 term_mismatch=1300 mutant=divergent
#   selected_rows=111 selected_terms=527 selected_simple=true
#   guard=sufficient_not_necessary coefficient_unchanged=true
# PASS_REGISTERED_RESEARCH_PACKAGE research/scatter-parity-contract-audit outcome=completed

Local replay passes on Linux with CPython 3.12.13, 3.13.5 and 3.14.4, including -I -O -Werror -B; metadata/pin check passes. Hosted CI is queued.

Limits

Finite, source-bound audit: no native backend, no full cold replay, no formal-kernel result, no independent human review. It does not claim the mutant invalidates the selected construction, and it does not adopt or modify any guard. simple-scatter-v1 remains a valid, producer-compatible choice; the finding is that uniqueness is a producer restriction rather than the exact semantic condition, with [0,0,0] as the smallest agreeing row it rejects.

Credit

Counterexample, prior guard and boundary handoff: Nick Petalas/PR2977 and discussion #2976. Opt-in enforcement at both entry points: Arham Tahir/PR2992. Exporter: Dylan Crow; selected word and checker retain the eumemic / DreamingOfClouds / PR589 lineage. This package adds the exact characterisation as a checked domain statement, the guard-versus-agreement comparison, the producer row-space check and the term-count observation. Apache-2.0.

Prepared by @kumagi with substantial AI assistance (DeepSeek V4.1 Flash in the OpenCode harness). No independent human review claimed.