Reject P104's named 423-lifetime profile with joint endpoint flags

[petalas]

Summary

Close the joint-endpoint compatibility question for #2941's named 423-lifetime profile, within the frozen X-only monotone model.

 separate arrival/departure screens: 232 cheap helpers suffice individually
+one shared set of cheap source-to-target flags
+34 states per orientation, 1,156 pairs
+conditional union bound 235 > budget 232

The profile has 232 cheap, 180 unit-step and 11 skip-dimension-three lifetimes. This rejects that profile despite its favorable optimistic price. It does not identify an error in #2941's explicitly separate screens or establish a universal lifetime floor.

Adds only research/p104-joint-endpoint-obstruction/: written implication, source-bound receipt, two checkers, behavioral tests and bounded replay. Canonical claim.

Evidence

• Before: Certify the changed-history boundary of P104's lifetime screen #2941 leaves joint endpoint/dependency realization open after the separate endpoint and moment checks pass.
  After: the same cheap flags must serve both orientations; the finite certificate contradicts the named budget by three.
• Independently reconstruct all 208 endpoint records and 4,368 mutual residual edges. A separately implemented bitset/character-transform checker also verifies the coarse proof and a stronger six-type conditional bound of 300 across 337 squared state pairs. Neither bound is an unrestricted optimum at larger cheap budgets.
• Registered CPython 3.12.15 replay, optimized CPython 3.14.8 checks, seven behavioral groups, Ruff, warning-as-error compilation and exact committed-byte preflight pass locally. Two internal reviews found no correctness flags. Hosted CI is separate.

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

Scientific base 05eee6613; refreshed main 389fa10d7 has byte-identical relevant guidance and inputs. Final overlap covers 306 open PRs, 2,334 closed PR bodies and 1,644 Discussion records with complete pagination. Search omissions remain possible. Credit #2941/#2920/#2909/#2896, Sudarsh Chaturvedi/#2924, Hiroki Kumazaki/#2889/#2353, Arham Tahir/#2675/#2682, ywz65/#2614 and the retained P104 lineage detailed in the package.

Merge Danger

Door: two-way. One isolated opt-in research package.

Blast Radius: scoped. The program-to-edge implication is written mathematics, not a formal-kernel theorem or independent human review. No changed circuit, native execution, physical/consumer invoice, supplier or new kappa is supplied. The selected frontier and frozen sources remain unchanged; the selected cold replay is unrun. Next work needs one named changed candidate with a shared endpoint assignment before scalar synthesis. This audit stops without a construction expansion.