A pipeline test pins a generated fetcher body with code strings

SingleRecordPayloadPipelineTest’s direct-@table` two-step case asserts on the text of a generated fetcher body: it counts transactionResult( occurrences, does body.indexOf(rowsName
"(keys, env)") arithmetic against the transaction call site, and pins .doesNotContain(".select(") on the fetcher beside .contains(".select(") on the reentry companion. docs/architecture/principles/development-principles.adoc bans code-string assertions on generated method bodies at every tier ("they test implementation, not behaviour, and break on every refactor; the compile and execution tiers replace them"), and notes the ban is review-enforced rather than build-enforced, which is how these survived. They predate the ban’s current phrasing: git blame puts them on R75 (the direct-@table two-step emit and its durability pins) and R314 slice 4 (the named DML reentry unit), with R482 passing over them for a different reason.

What they are reaching for is a real invariant and worth keeping: the write commits before any follow-up read can fail, which currently survives as call ordering inside one emitted body ; the rows<Name>(keys, env) call sits after the transactionResult(...) call site, and the companion, not the fetcher, owns the .select(...). The task is to restate that at a tier that can see it rather than to delete the coverage: the durability claim is an execution-tier fact (a committed write survives a failing follow-up read), and "which method owns the SELECT" is a structural fact about the emitted TypeSpec that can be asserted over MethodSpec identity and call-graph shape instead of over rendered source text. R885’s Done gate set the precedent for the second half, narrowing two such pins in TypeFetcherGeneratorTest into one explaining helper rather than multiplying them; R880’s set the precedent for the first, deleting a FetcherPipelineTest code-string pin and restating its claim at the execution tier.

Found during R687’s In Review → Done gate, in the same test class R687 extended. Not R687’s to fix: that item added no code-string assertions (its own pins are classification verdicts, rejection messages and a TableRef equality), and the delivered change does not touch these lines.