Meta-Extensible Context: Constructive language extension, temporal integrity, and native alignment in a finite developing kernel

A developing context system must distinguish a limitation of its present representation, a limitation of its candidate language, and a lack of evidence. It must also be able to refuse its own proposed upgrades. We investigate these requirements in a finite relational environment with protected history, versioned evidence, separate qualification, continuing duties and a fixed alignment reference. The original persistent campaign rejects all 15 deliberately inadmissible proposals, including self-authorisation, while expanding warranted decisive coverage from 144 to 166 to 180 of 1,530 support/question cases. Only 180 cases permit a decisive answer. A subsequent branch chooses among eight supplied generator revisions and withholds extension when information is absent. A further constructive experiment starts from variables and a supplied NAND primitive, generates all 256 Boolean functions, and qualifies a useful composed feature without receiving parity as a finished candidate. The first useful tree is sound but unaffordable; a shared-expression search finds an admissible implementation under the declared cost limit. Ten matched governance knockouts expose failures in nine cases and redundancy in one. A mutable-history preflight defect, then reproduced by intervention, establishes a concrete temporal-accountability vulnerability even when answers remain correct. These are finite, reproducible integration results, not novel logical primitives, independent replication, economic superiority or general alignment. The paper separates installed-ecology gain from generator reachability gain, supplies conditional soundness arguments, and preserves failures and raw evidence in six portable campaign capsules.