Tooling: exact DAG union propagation speeds historical tensor audit 1.44x

[steph4n-gh]

The historical disjoint-tensor frame audit repeatedly enumerates reachable targets to compute their ground-coordinate unions. This opt-in adapter propagates the same exact unions through the unchanged reverse DAG traversal. Tooling improvement; no new κ.

Five alternating fresh-process pairs on macOS arm64 / CPython 3.14.4 reduce median complete-audit wall time from 5.014369 s to 3.492928 s (1.435577×). All ten runs reproduce the full original certificate byte for byte, including its original source bindings. Raw samples and source/code hashes are included.

The adapter preserves all three ordered maps for 60,372 h26 active nodes and their 9,206,600 reachable-target memberships. Full descendant bitsets, source unions, rectangle guards, both physical orientations, coefficient/embedding checks, small exact residual bases, invoices and rational inequalities remain checked. Only research/disjoint-frame-unions/ is added; the default verifier, selected frontier and immutable historical sources remain unchanged.

Reproduce from the repository root:

python3 -I -B research/disjoint-frame-unions/verify.py

Expected: PASS_DISJOINT_FRAME_UNIONS, complete historical audit, receipt SHA256 3f2b37542f9581f9d3238cccfce19fd3e2964c89ec4af4ff047fb8041f5c3cd1.

Validation:

• Eight differential/independent graph tests pass normally and under isolated -O, including a separate BFS oracle on 88 seeded DAGs, all h26 map entries/order, aliases, repeated children, output permutations, gapped active nodes, and the five original historical tests under the adapter.
• Five replay controls pass: source/adapter corruption, missing source, changed or incomplete receipt, patch restoration after failure, optimized-full-replay rejection and existing-output preservation. Full inherited audit correctly refuses -O.
• PYTHONDONTWRITEBYTECODE=1 make verify-tests: 661 tests in 96 isolated modules and 20 patch checks pass. Selected-record and whitespace checks pass.
• Minimal offline relocation with spaces in its path reproduces the full receipt. All 40 inherited pins match the base and refreshed main; no default certificate was regenerated.

Base: 2921874911956718fcaaaf9396fe8267b3cfa9f0. OS-random assignment seed 88, verification engineering. Local claim, candidate-selection rejection evidence, overlap review and full validation limits are included. Related #863/#981 change root-contraction constructions; #950 independently considered this traversal and chose central-factor packing. New #1029/#1041 optimize U8 phase recipes. None supplies this frame-union optimization.

Credits Douglas R. Colkitt's original construction/verifier and Ross Richardson's #950 opt-in replay/benchmark pattern. Prepared by Stephan Arrington (@steph4n-gh) with substantial OpenAI Codex assistance, including separately tasked AI graph checks and review; Apache-2.0. No independent human review is claimed.

Full make verify, selected-supplier cold replay, Lean and hosted Linux CI were not rerun. This historical checker does not accelerate the selected T37 replay or establish a new bound; all inherited all-size/compiler, realization, hosting, precision/tape, analytic and bit-interface obligations remain.

[Mister-Meeseeks]

100-PR batch review at 8c5dfb29368481f6bb7bb210ba08d8f58d9f4ef8.

Merged to main, preserving the original contributor commits and package attribution.

Reviewed scope: Opt-in disjoint-tensor reachable-target union propagation with complete historical certificate identity; no new construction.

The pinned scoped command passed; the batch ledger contains its command, input hashes, environment and log. This incorporates the research/tooling contribution and does not change the selected κ or discharge the inherited multiplication-theorem assumptions.

Batch review, attribution and evidence. Thank you for the contribution.