← board

Track guessed as N 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.

This expectation records a REFUSAL (ValueError). Before treating a converged bisect range as an accusation, check whether the named commit IMPLEMENTED the thing being refused -- a feature landing makes its own refusal test go red, and the bisect converges on it correctly. Not a verdict; the tool cannot decide this one.

origin/master has advanced 47 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-nilpy#src:test/test_nilpy_max_min_iterables.npy red at 0200df7eabcd (auto-filed by twatch)

Repro

tools/testmgr.py --tier full --job 'test-nilpy#src:test/test_nilpy_max_min_iterables.npy' at 0200df7eabcd33796c0f7ac151b80aafbf75b5fb

Range

The named sha 0200df7eabcd 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 0200df7eabcd, last good 3f854c927aac, 6 commit(s) in range — the watcher narrows this by idle bisect; check tstate/TSTATE.md for the current range.

Log tail

ok: /tmp/testmgr-scratch-3750893/test_nilpy_maxmin_iter26  [code=1272141B  data=55351B  bss=53196B  procs=1860]
Unhandled exception: TypeError: expected a number, got object
--- test/test_nilpy_max_min_iterables.expected	2026-08-29 16:03:42.850941360 +0000
+++ -	2026-08-30 02:08:28.243527547 +0000
@@ -3,13 +3,3 @@
 3 1
 z a
 none none
-b b
-3 1
-c a
-3 0
-99 97
-6 2
-9 2
-[1, 2] [3]
-7 3 2.5 a
-ValueError: max() iterable argument is empty

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

Resolution (frankA, 2026-08-30)

Cause: 7b73a385d (feat(N): list.sort(key=) — the callable→Pointer coercion), inside the watcher's range and confirmed by bisect — 7b73a385d^ prints b, 7b73a385d raises. Both builds seeded from their own committed stable and fixedpoint-verified; see the seeding note below, which is why the first two bisect attempts were void.

That commit moved the callable→Pointer coercion into PyBindKwArgs, i.e. into the KEYWORD path. The coercion runs AFTER overload selection, so it made the key: Pointer candidates viable for a keyword call — and therefore newly reachable for receivers that cannot be walked as a list. A dict handle bound to max(l: TPyList; key: Pointer) and was dereferenced as a TPyList: l.count/l.at read a TPyDict's fields, and the garbage surfaced far away as TypeError: expected a number, got object. Before the commit these calls fell through to the variant arm, which normalises through pylist_v.

The measured boundary is what named the mechanism: max(d, f) positionally works and max(d, key=f) does not, and on the failing path the key function is never called.

Fix — complete the receiver set rather than special-case the dict. sorted has carried TPyDict and AnsiString overloads since bug-nilpy-sorted-over-a-string-segfaults, whose comment records this exact failure ("the AnsiString handle bound to the TPyList overload and was dereferenced as an object"). min/max had only TPyList/TPyIter/TPyRange. Added the two missing receivers, delegating to the TPyList routine (d.keylist, pystr_charlist(s)) so the empty-sequence ValueError and the first-wins tie rule stay in one place. No frontend file touched.

A Variant keyed pair was written and then REMOVED — it bought nothing. I first read "tuple/generator now work" as its effect; that compared a literal receiver before against a named receiver after, changing two axes at once. With the pair deleted the whole suite still matches, so it is not in the change: an overload that alters how a variant unwraps is exactly what the decl-order comment above min(it: TPyIter…) warns about, and it has to earn its place.

Gate: test_nilpy_max_min_iterables matches its CPython-generated .expected; the eight neighbouring sort/sorted tests still match; the shapes the decl-order comment says broke historically (min(v)/sum(v) over a variant list, 2/3/4-arg numeric, default=) all match CPython. make compiler/pascal26 converged. Baseline re-run with pyeval.pas stashed fails — and note the compiler binary is byte-identical with and without this change (1ff8acbe123b), because compiler/builtin/** is consumed when compiling a .npy, not linked into the compiler, which is what makes that baseline cheap and honest.

The new rows needed their own baseline. With the whole file, the baseline dies on line 6 (the original row) and never reaches rows 16-20, so the file cannot show they are load-bearing. Run alone against the baseline library: dict and named-str fail, and tuple/generator/list-of-lists already passed — they are pinned as guards, not claimed as fixes.

Residual, filed separately — NOT closed by this

A literal str receiver with key= is still broken and is a regression from the same commit (max("bca", key=f) printed b at 7b73a385d^). A named str works. The keyword promoter re-targets on the argument's STATIC type, and a literal reaches no keyed overload, so this needs pyparser.inc, not the library. A literal tuple/generator fails the same way but is pre-existing (broken at 7b73a385d^ too) and is bug-nilpy-keyword-arg-vs-overload-set. → regression-nilpy-a-literal-str-receiver-with-key-reaches-no-keyed-overload

Not a defect: max(d, f) positionally answers b where CPython raises. NilPy is upward-compatible one direction only; accepting what CPython rejects is a divergence, not a bug (CLAUDE.md, Track N).

Seeding note for the next bisect in this repo

make compiler/pascal26 seeds from ./compiler/pascal26 — the binary already in the tree. Walking backwards, that is the previously tested commit's binary, and an older one fails with undefined variable (__pxxblockmove). Seed each commit from its OWN stable_linux_amd64/default/stable_pinned, rm the fixedpoint stamp, touch the sources (a copied-in seed is newer than them — CLAUDE.md's no-op trap), and require converged after in the log before accepting a verdict. A plain git bisect start here also spans ~5600 commits, nearly all watcher tstate publishes; path-limit it or test candidates directly.