Reject the P104 253/147 profile by endpoint demand saturation

[petalas]

Summary

Close PR #2896's endpoint-assignment handoff for its surviving 400-lifetime profile in the frozen X-only monotone P104 model.

253 cheap + 147 unit-step helpers
aggregate support     4663 >= 4656   passes
endpoint saturation   3972 <  4656   rejects, deficit 684

The written necessary bound caps useful support at each endpoint's demand. For 400 helpers restricted to these two history types, at least 192 unit-step helpers are necessary. This threshold is exact only for the separate endpoint-count relaxations; it supplies no geometric or chronological correction circuit.

Adds only research/p104-endpoint-saturation/, with proof, original-input reconstruction, exact greedy and independent integer-DP checks, source pins and registered replay. Credit Nick Petalas/#2896, Hiroki Kumazaki/#2889/#2353, ywz65/#2614, Arham Tahir/#2675/#2682, Sagar Thakrar, Philippe Van Velzen and the retained geometry/native lineage. Full attribution and scope are in the README and TASK.

Canonical claim and handoff. Reviewed commit: 4f394cfd3244b336e59b7adaded883277b66b637.

Evidence

• Before: Bound P104 endpoint support and reject the 406-lifetime profile #2896's 253/147 profile passes its aggregate support rows and favorable cost relaxation. After: the new exact allocation screen rejects it in each orientation separately.
• All 208 endpoint records are reconstructed from the original pinned PR1972/1991 inputs and match the exact Bound P104 endpoint support and reject the 406-lifetime profile #2896 snapshot. Greedy and DP curves agree at every one of 209 unit budgets per orientation.
• Registered normal/optimized CPython 3.12.15 replay and five behavioral groups pass, including 780 exhaustive toy allocations, a saturation-removal mutant, Git-free execution and source/typed-receipt corruption controls.
• Ruff, warning-as-error compilation, selected-result binding and committed-byte preflight pass. Bounded internal review found no substantive defect; this is author-checked evidence, with shared geometric premises.

python3.12 -B scripts/research_package.py --package research/p104-endpoint-saturation

Started from canonical main 622853501; refreshed main 11720ce76 has unchanged relevant guidance, inputs, selected record and replay runner. Final overlap covers 488 open PRs, 2,114 closed PR bodies and 1,581 Discussion records with complete pagination. Unpublished work and arbitrary diffs remain outside coverage. Hosted CI is separate from the local checks.

Merge Danger

Door: two-way. One new opt-in package; predecessors and the selected frontier stay unchanged.

Blast Radius: scoped. This rejects the stated two-type profile under the inherited X-only monotone contract. Actual subspace unions, joint assignments, chronology, restoration, hosting, precision and a complete consumer cost remain open. No general P104 impossibility, supplier admission, new kappa or full theorem verification is claimed.