← board

Track guessed as P from the test source. The ranker reads frontmatter, so this line — not the body — decides who works it; correct it if the guess is wrong.

origin/master has advanced 9 commit(s) since this sha. Re-verify at current HEAD before acting — the callback is tagged to the sha that was tested, which may no longer be the state of the tree.

regression: test-asm#src:test/test_asm_emit_rv32.pas red at 108ac182bed6 (auto-filed by twatch)

Repro

tools/testmgr.py --tier native --job 'test-asm#src:test/test_asm_emit_rv32.pas' at 108ac182bed6cd40244fd24108d6664d6cf1b2f0

Range

The named sha 108ac182bed6 CANNOT be the cause — it touches no buildable file (docs / tickets / tstate only). It is the sha that was TESTED, i.e. the upper bound of an untested range; the cause is somewhere below it.

bad 108ac182bed6, last good c951ec710b33, 2 commit(s) in range — the watcher narrows this by idle bisect; check tstate/TSTATE.md for the current range.

Log tail

pascal26:43: error: undefined variable (AIntToStr)
pascal26:44: error: expected comma or close parenthesis
(tail)
pascal26:43: error: undefined variable (AIntToStr)
  in: compiler/rv32enc.inc
  near:  what   displacement   AIntToStr >>>  v  
pascal26:44: error: expected comma or close parenthesis
  in: compiler/rv32enc.inc
  near: v    is outside the encodable range   AIntToStr >>>  lo

Stub ticket: signal only. Track T agent (face 2) enriches or a dev track takes it from the repro line.

2026-08-30 (coordinator) — RETRACKED P → A, on the error's own text

The auto-file guessed P from the test path, and the banner in this ticket says plainly that the guess is what the ranker reads. The guess is wrong, and the failing line says so without needing a repro:

pascal26:43: error: undefined variable (AIntToStr)
  in: compiler/rv32enc.inc

AIntToStr is defined in compiler/util.inc, pulled in at compiler/compiler.pas:64 — "shared helpers owned by no frontend/backend". So this is an include-order fault in the compiler's own build, in a reduced configuration: rv32enc.inc calls a helper that is not in the stream yet. Nothing about it is the Pascal frontend. Lane A, S-flavoured (riscv32 encoder); related to feature-a-build-a-reduced-compiler-by-selecting-frontends-and-targets, which already names this test.

Why the retrack matters more than the fix here. At track: P and prio: 70 this sat near the head of a queue whose agents own pasparser_*.inc and would have opened it, found compiler/rv32enc.inc, and handed it back — while the lane that owns the file never saw it. A guessed track does not read as a guess in ready output; it reads as a priced decision. Retracked on the error text, not on a hunch; if the guess was right after all, the evidence to overturn this is one grep.

Not re-verified at HEAD by me. Track T reports it STILL-RED on every native report from 06:57Z to 08:45Z, so it is standing rather than flaky, but the attributing sha 108ac182bed6 is itself a tstate-ticket(...) commit — a docs-only change that cannot have caused this. A tstate red attributes to the sha it swept, not the one that caused it. Whoever takes this bisects; do not start from that sha.

2026-08-30 (frankA, Track A) — RESOLVED: the harness's mock environment, not the compiler's build

Reproduced at HEAD (4413c141a, self-hosted binary aa78a7faf63a, converged after 1 round(s)):

./compiler/pascal26 -Fucompiler test/test_asm_emit_rv32.pas /tmp/.../rv32emit
pascal26:43: error: undefined variable (AIntToStr)
  in: compiler/rv32enc.inc

It is not an include-order fault in the compiler, which is what the error text suggests and what the coordinator's read proposed. make compiler/pascal26 converges at this sha: util.inc is pulled in at compiler.pas:64 and rv32enc.inc at :81, so in the real build AIntToStr is in scope 17 includes before its call site. The compiler was never broken.

The failing program is test/test_asm_emit_rv32.pas, a standalone oracle harness that mocks the byte sink (EmitB/EmitI32/Patch32/Error) and then {$include}s the real shipped compiler/rv32enc.inc to assert its bytes against llvm-mc. Its mock environment is hand-rolled, so every helper an included compiler file grows must be mocked by hand too — and 2f81d8008 gave rv32enc.inc a RISCVRelCheck whose Error() text formats the offset with AIntToStr. The bisect is exactly right; the bad commit is a good fix that outgrew a mock.

Fix: an AIntToStr mock next to the AppendChar mock already in that file (Result := IntToStr(n); the harness uses SysUtils, and exact over the Integer range). Including util.inc for real is not available to the harness — that file opens on AppendChar, which lives in lexer.inc, so it drags the lexer behind it, which is why AppendChar is mocked by hand there in the first place.

The harness is green and the guard is on a live path, not dead code — EmitBType calls RISCVRelCheck unconditionally, and the run reaches PASS: jal forward / PASS: beq backward before ALL RISC-V ASM EMIT TESTS PASSED.

Sibling sweep (root-cause-over-microfix): rv32enc.inc is the only exposure

The class here is a harness-included compiler file calls a shared helper the mock environment lacks, and this file has now met it three times — AsmRv32ProcessInlineLine, InlineAsmLineHoleN (both 2026-08-21, both recorded in its own comments), and now AIntToStr. So the sweep is over the whole class, not over riscv32:

So no sibling harness is broken today, and no other harness needs a speculative mock. xtensaenc.inc's XtensaRelCheck — the template RISCVRelCheck was mirrored from — uses AIntToStr the same way and is fine for the dull reason that no harness includes it: test_asmcore_xtensa.pas links lib/asmcore rather than {$include}ing the compiler's encoder.

What is left standing, deliberately. The five test_asm_emit_*.pas harnesses duplicate their mock preludes (4 of the 5 carry their own copy of AsmTextTrim), so the concept "the compiler's shared helpers, mocked for an encoder harness" is served by five hand-rolled copies — the shape devdocs/dev/normalise-dont-special-case.md warns about. It is not folded into a shared prelude here: this is a live red at prio 70, the refactor puts five currently-green harnesses at risk to close one ticket, and the rot is no longer silent — chore-a-sweep-the-unwired-tests-into-the-suite wired all five into make test-asm, which is why this surfaced the day the encoder changed instead of years later. Filed as [[idea-a-fold-the-asm-emit-harness-mock-preludes-into-one]].

Gate: make compiler/pascal26 (converged, 1 round) + the repro compiles and prints ALL RISC-V ASM EMIT TESTS PASSED.