[flyingviper]

PR5569 leaves owned-temporary cleanup as an open route to a cheaper P104 invoice. This package proves scoped cost floors and delivery obstructions for its exact word. No new kappa or global optimality claim.

The distinct results are:

• Seven fixed-boundary owned delivery blocks already attain their minimum raw recursive paid moment, for every 0<t<=1, even allowing arbitrary rational internal words and zero-initialized private auxiliaries. Outside side information, changed productive targets and work shared across block boundaries are excluded.
• A global loss-space invariant gives an H7 floor for preparing one unchanged producer value at its retained delivery line. The eight individual bounds are explicitly not additive.
• A joint rank-eight cut requires at least eight owned crossings for the eight unchanged observations, even with sequential delivery and shared preparation. The source attains the count. This does not imply an 8H7 tariff: crossing ranks may change.
• Owned738's exceptional multi-level route cannot lose a stop by retiming its five deliveries within fixed neighboring contacts. A unique-public-basis certificate also excludes a public multi-partner replacement of one nominated delivery at every 29,688 recipient/cut pair, over arbitrary rational coefficients and every possible common binary frame.

The independent 5,252-case public-window census is retained as corroborating source validation. Submission refresh found that Ali Romman had already published the same domain and separator mechanism in PR5579; we do not claim that overlap as novel. Discussion5575's compensated source substitution changes dependencies and remains outside these obstructions. PR5580's new contact gain does not contradict them; our finite certificates are pinned to PR5569, not automatically asserted for its descendant.

Review starts at research/p104-owned-cleanup-obstruction/README.md; the separate PROOF-*.md files state each model, proof and limitation. RESULT.json contains exact covectors, nonzero minors, boundary ranks, window examples, witness hashes and semantic controls. Further progress must change the excluded dependencies/boundaries or produce a fully charged shared itinerary. Merely adding partners or coefficients to the tested common-chart mechanisms cannot help.

Validation on Windows / CPython 3.12.14:

python3.12 -I -B research/p104-owned-cleanup-obstruction/verify.py
python3.12 -I -O -B research/p104-owned-cleanup-obstruction/verify.py
python3.12 -B scripts/research_package.py --package research/p104-owned-cleanup-obstruction
python3.12 -B scripts/check_research_commit.py --package research/p104-owned-cleanup-obstruction --check-source-history

All pass. Direct replay takes about four seconds, uses only the standard library, checks all 743 source scalar endpoints, and includes positive fixtures, 15 corruption controls and additional omitted-delivery/scope controls. An independent agent implemented the window census separately; independent agents reviewed the new proofs, boundary ranks and inverse-covector argument. These are author checks, not independent human acceptance. The official committed export also passes isolated optimized replay with read-only files, empty PATH, no Git, and unchanged hashes. The original PR5569 complete native/caller/funding verifier passed unchanged before this analysis; this small package does not repeat or claim a new complete bound. Hosted CI is pending publication.

Base canonical main: 5e0052ca0aa3036a3afa8b0438927377c8d97d53. Source: PR5569 head af9f091ac6f30c6eaacb9c27777bfbe2d0e5ffdf. Two unchanged compressed words are retained with hashes and attribution in a first commit; standard source pins refer to that ancestor in this publication's own history. This avoids relying on unmerged commits present only in a local object cache. Submission overlap coverage was the 100 most recently updated open PRs plus relevant public Discussion pages, not an exhaustive novelty claim.

Prepared for flyingviper with substantial OpenAI Codex assistance and independent research/review agents. Credit Joel Pulikkan/GamingPuzzled (#5569, #5472), DreamingOfClouds (#5561, #5553), thu-nmrc (#5551), Ali Romman (#5544/#5545 and concurrent #5579), and the retained Gareth Davidson, David Arias, Cinna Davis, zxv, Cao Yifan and earlier P104 lineage. Primary literature consulted is documented separately, without importing its cost model. Apache-2.0. Selected frontier and predecessor files remain unchanged.