What is being asked

Drop w2stress.pas (written for c93292fe4, currently in the W2 scratchpad) into the test corpus as an ordinary standalone test program. Nothing else is needed: tools/optdiff.sh globs every standalone .pas/.c in the corpus, shards it 12 ways, and compares stdout+stderr+exit code at -O0, -O1, -O2 and -O3. Once the file is in, it is swept at every level on every opt phase, forever, with no sidecar table and no expected-output file to maintain.

This is Track O's own program and Track O's file to place, which is why it is filed here rather than done by T.

Why it is worth doing

optdiff covers the corpus broadly but not densely in the constructs the register-pressure campaign actually changed. w2stress.pas is deliberately dense in exactly those: all five in-place ALU ops at every integer width driven past their wrap point, signed and unsigned narrowing, self-referencing assignment, non-commutative x := x - y, {$Q+}, in-place stores inside try/except, var and value params, pointer arithmetic.

The campaign's own near-miss is the argument. frank-optimize-b4's first W2 build silently refused the hottest shape in the language — it guarded on IR node types, and a for loop's own increment carries tyUnknown, so it fired on s := s + j and not on i := i + 1. Half the win, every output byte-identical, every test green, caught only by disassembling. A dense program in the sweep does not catch a missing optimisation either — nothing in a correctness suite can — but it is where a wrong one would show up first.

Notes for whoever places it

• The {$Q+} trap does not apply here. The program's FPC-oracle problem — pxx detects a LongInt overflow that fpc -O2 misses, recorded in devdocs/dev/pascal-dialect-divergences.md — only bites a differential harness with an external oracle. optdiff has none: the other levels are the oracle. That is the reason to prefer this route over an FPC differential, and it is Track O's own reasoning from the originating ticket.
• If any level legitimately diverges (nondeterministic output, timing), the sweep has tools/optdiff.skip for named patterns with a reason — but a real divergence here is a codegen bug and should be filed, not skipped.
• It must be standalone-runnable and terminate well inside the 30s base timeout, or the sweep records it as a skip.

Provenance

Resolved out of chore-t-nothing-in-the-matrix-runs-o3-so-no-failures-is-unfalsifiable, whose central premise (that nothing in the matrix compiles at -O3) was refuted by measurement: opt has run 701 times, has gone non-GREEN 30 times, and four -O3 bugs in done/ came from it. Two of that ticket's three real items were done by Track T; this is the third, and it belongs to O.

Resolution — 2026-08-28: already done, in d2eafe1e3

The program landed in the corpus with item 3, before this ticket was filed, as test/test_o3_resident_inplace.pas (renamed from w2stress.pas, and extended past the version this ticket describes — it also carries the unwind-through cases written to attack item 3's exception-frame gate: residents in a body with no handler of its own, unwound by a raising callee).

Verified against this ticket's own three requirements rather than assumed:

• Standalone-runnable at every level — builds at -O0, -O1, -O2, -O3.
• Well inside the 30s timeout — 0.00s at -O0, the slowest.
• All four levels agree — identical stdout+stderr, which is exactly the comparison optdiff.sh makes.
• Not matched by any pattern in tools/optdiff.skip, so it is swept, not skipped.

optdiff.sh globs the corpus by extension, so no registration step exists and none was missed. It is swept at every level on every opt phase from here on.

A .expected file was added alongside it. optdiff does not need one — the other levels are its oracle, which is this ticket's own reasoning for preferring that route — but the ordinary test suite does use it, so the program is covered twice by different mechanisms.

Process note for the next person appending to a ticket like this one: the no-full-suite hook pattern-matches the COMMAND TEXT, so writing this section via a shell heredoc was refused — the prose quoted the corpus glob literally and the hook could not tell documentation from a regression run. It was right to refuse on what it could see. Write ticket prose with a file tool, or describe the glob rather than spelling it.

Log

• 2026-08-28 — resolved, commit 9e9e84ab5.