Re-priced 15 -> 35 by the coordinator, 2026-08-30. frankA's call to file this rather than do it during the p70 red was right, and the reasoning for 15 was sound at the time it was written. What moves it is the count, not the risk: this is the THIRD instance in the same file (AsmRv32ProcessInlineLine, InlineAsmLineHoleN, now AIntToStr), and devdocs/dev/root-cause-over-microfix.md prices two as a smell and three as a design flaw. It also has the one property a low-prio ticket must not have: it re-arms itself. Every encoder change from here can turn a green harness red for a reason that is not in the encoder, and the five harnesses were only recently wired into test-asm, so the population that can trip it just grew.

Not raised further, for frankA's own reasons, which stand: the rot is no longer silent, the refactor risks five green harnesses to close nothing currently red, and the preludes are not identical so it is not mechanical. 35 says "do it deliberately, on a quiet lane, before the next encoder campaign" — not "do it now".

Fold the five asm-emit harnesses' mock preludes into one shared include

The test/test_asm_emit_{x64,386,a64,arm32,rv32}.pas oracle harnesses each mock the compiler environment by hand — a byte sink (EmitB/EmitI32/Patch32), Error, CaseEqual, AppendChar, an empty inline-asm line pool, the AsmText* string helpers — and then {$include} the real shipped encoder to assert its bytes against llvm-mc. The mocks are duplicated: 4 of the 5 carry their own copy of AsmTextTrim, and each carries its own copy of most of the rest.

The failure mode this creates is that any shared helper an included compiler file grows must be added to each prelude by hand, and nothing connects the two edits. Recorded instances, all in test_asm_emit_rv32.pas alone:

The last one is [[regression-test-asm-test-asm-emit-rv32]] — a good encoder fix (2f81d8008, PC-relative range guards) that reddened the suite purely by calling a helper the mock did not have. This is the shape devdocs/dev/normalise-dont-special-case.md names: one concept ("the compiler's shared helpers, mocked for an encoder harness") served by five hand-rolled mechanisms, so a fix to one arm leaves four.

Why it is prio 15 and not higher

The rot is no longer silent, which is most of the value already. chore-a-sweep-the-unwired-tests-into-the-suite wired all five into make test-asm, so 2026-08-30's instance surfaced the day the encoder changed rather than years later — that is the difference between the first two rows of the table and the third. What is left is the per-instance cost of adding one mock in one file, which is small.

Against that, the refactor puts five currently-green harnesses at risk at once, and it is not purely mechanical: the preludes are not identical (386 carries 26 mocked entry points to rv32's 19, and the byte sinks differ), so a shared test/asmharness_prelude.inc has to separate the genuinely arch-independent part (Error, CaseEqual, AppendChar, AIntToStr, AsmText*) from the per-arch part, which is the actual design work.

Shape, when someone takes it

• test/asmharness_prelude.inc = the arch-independent mocks, including every compiler/util.inc export (AIntToStr, RealTypeKind, TargetIsEspClass, SoftFloatMissing) whether or not an encoder calls it today — the point is that the next one that does costs nothing.
• Each harness keeps its own sink and its own arch include.
• Gate: all five compile and print their ALL <arch> ASM EMIT TESTS PASSED line (they are wired into make test-asm).

Note that util.inc itself cannot simply be included by a harness: it opens on AppendChar, which lives in lexer.inc, so pulling it in drags the lexer behind it. That is why AppendChar is hand-mocked in these files today, and it is the constraint the shared prelude has to live with.