← board

The struct module

import struct reaches no unit and no shim.

Why it is next

mimic_array landed 2026-09-09 and cleared the array wall in world.py; the module's next error is this one. That is the honest shape of the lekkerzeilen blocker census — see the umbrella — and it is why this ticket exists rather than the array one simply being reopened.

CORRECTION 2026-09-10 — the population, re-measured at HEAD

The call sites below are labelled rd.py and that is wrong; they are capture.py's. rd.py imports math and not struct, and compiles clean at 786a00f9a. Re-measured first wall per module, compiler 61f8a78f8aae:

module first wall
world struct
capture ctypes (then zlib, then struct)
rd none — compiles clean
gauges, app queue

grep -l '^import struct' over the runtime package returns exactly capture.py world.py. So the first-wall count for this ticket is 1, not 3, and clearing it is not measured as sufficient for capture — which needs two other things first, one of them deliberately out of scope.

Recorded rather than quietly edited: the wrong number was in the summary, which is the part everyone reads, and it would have ranked this above the queue ticket on a population that does not exist.

Measured surface (lekkerzeilen, 2026-09-09; file labels corrected above)

struct.pack(">I", len(payload))                       # rd.py
struct.pack(">IIBBBBB", width, height, 8, 2, 0, 0, 0) # rd.py
struct.pack("<h", 1) != struct.pack("=h", 1)          # world.py — endianness probe
struct.pack("<f", 1.0) != struct.pack("=f", 1.0)      # world.py
struct.unpack("<%df" % (len(verts) // 4), verts)      # scenery-side, RUN-TIME format
struct.unpack("<%dI" % (len(indices) // 4), indices)

So: byte-order prefixes < > =, formats I i h H B f d, and a repeat count that is computed at run time — the format string is not a literal, so a frontend-side compile-time format parser is not an option; the shim parses the format at run time, as CPython does.

unpack returns a TUPLE, and both call sites wrap it in list(...).

Shape

A .pas shim (lib/rtl/mimic_struct.pas), for the reason mimic_array states in its own header: the job is byte reinterpretation, which Pascal does with a pointer cast and pure Python can only do by hand-rolling IEEE-754. The typecode size/signedness table is the one mimic_array.SetTc already carries; the two should share it rather than growing a second copy.

The = vs < distinction is load-bearing and is not cosmetic: world.py's only use of struct is to ask whether the host is little-endian, by comparing the two encodings. A shim that treated = as < would answer "never swap" on a big-endian host and decode every terrain file backwards — silently, and only there. Whatever else is subset, that comparison has to be real.

RESOLVED 2026-09-10 (frankB) — lib/rtl/mimic_struct.pas

Byte-identical to CPython across all 33 rows of test/test_nilpy_the_struct_module.npy, whose .expected is generated by CPython and not by us — including the float32 precision loss (0.10000000149011612), the run-time repeat count, and the tuple repr.

The typecode table is SHARED, not copied. mimic_array's twelve rows moved out of array_.SetTc into a unit-level PyTypecode exported from that unit, and this shim uses it. Our fixed widths ARE struct's standard sizes — l is 4 bytes here and 4 under any byte-order prefix in CPython — so one table really does serve both. Copying it would have put mimic_array's divergence note into two files that must agree with no way to notice when they stop. One shared table, two vocabularies: u is array's Py_UCS4 and is refused here.

= is resolved against the actual host order, probed at run time. Not a {$IFDEF ENDIAN_LITTLE}, which is decided by whichever compiler builds lib/rtl and is wrong the moment it is built once for a cross-target. world.py's only use of struct is the probe, so this is the one thing that could not be subset.

The tuple tag was free and worth taking. unpack returns a TPyList tagged PYSEQ_TUPLE, so print(struct.unpack(...)) gives (1234567,) and not [1234567]. Both corpus call sites wrap the result in list(...) and would never have noticed — which is exactly why it was worth one assignment: the next caller prints it, and a repr that quietly disagrees with CPython is found in someone else's output diff.

Stated divergences, neither tested against the oracle because a row asserting a known divergence is red by construction: bare l/L (ours 4 bytes, CPython native 8 on LP64) and @/no-prefix alignment (CPython pads, we do not). Both in the shim header.

pack is capped at 7 arguments — capture.py's ">IIBBBBB" is the widest call in the corpus, and a wider one is a clean compile error rather than a wrong answer. unpack has no such limit, and that is where world.py's run-time count lands. pack_list is public for a genuinely run-time-sized argument list.

Result on the target: cleared, and NOT sufficient

world.py's first wall moved from struct to math.atan2 (feature-nilpy-math-module-twelve-absent-names-measured, prio 75, already filed). capture.py is unchanged — its first wall is ctypes, settled as bind-natively. So this clears one module's first wall and compiles no module.

Log