← board

TRIAGED 2026-09-16 (frankuser). RE-LANED T -> A, and it is a REAL SILENT WRONG VALUE AT -O3. The Track T default was correct to be a fallback: the defect is not the harness. This is the uncovered -O3 ARM of [[bug-a-a-float-assigned-to-an-integer-lvalue-moves-the-bits-instead-of-converting]], which is in done/ — the general case was fixed and -O3 was never checked.

Minimal repro, and -O3 is the only level that is wrong:

program m;
function RetInt(F: Double): Integer;
begin
  Result := F;
end;
var v: Integer;
begin v := RetInt(4.7); WriteLn(v); end.
level output
-O0 5
-O1 5
-O2 5
-O3 -858993459

The full fixture prints got 4616977747989548237 want 5, and struct.unpack('<d', struct.pack('<Q', 4616977747989548237)) is 4.7 — the raw double bit pattern, not a number. So at -O3 the float->int conversion on a function RESULT is dropped and the bits are moved, exactly the parent bug's signature.

Why this is worth prio 80 and not 70: -O3 is on track for -O2 (CLAUDE.md), and this is a wrong VALUE with no diagnostic in an utterly ordinary construct. It does not crash; it returns a plausible-looking integer.

NOT a harness artefact — checked, because the two sibling shards WERE. frankb-56 closed optdiff shard2/shard10 on 2026-09-16 (311649be0): those compared argv[0], which differs per -O level directory, so nine days of red were a path string. I re-ran this row under the new OPTDIFF_FILES and then compared the program's own stdout directly at -O0 and -O3. It differs. test_c_gtk3_stock.pas (shard5) PASSES on a direct re-run and should be re-verified before anyone works it.

Track T by default: the FAILING STEP named no owner. Line 1 of 1 is tools/optdiff.sh --shard 0/12. The job's own src (tools/optdiff.sh, 1 file(s)) is NOT used here on purpose: it is what the job compiles, not what broke, and guessing a lane from it is what sent three reds in one job to the wrong lane. This is a FALLBACK, not a finding — nothing says the defect is Track T's. Re-lane it before working it.

origin/master has advanced 6 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: optdiff#shard0/12 at 285208414d3f in step 1/1, tools/optdiff.sh --shard 0/12 (auto-filed by twatch)

Repro

tools/testmgr.py --tier opt --job 'optdiff#shard0/12' at 285208414d3f332c0b13f2a084e671fbad114be7

Range

bad unknown, range unknown — there is no earlier passing sha to bound it, or the bound is not recorded. No idle bisect will happen; this one needs hand-triage.

Log tail

OPT DIFF -O3: test/test_double_to_integer_lvalue_rounds.pas (rc 0 vs 0)
optdiff skip SKIPLIST: test_rtti_field_get_by_name.pas
optdiff skip TIMEOUT-O0: devtest_tls_native_seam.pas test_emit_obj.pas test_halt_from_worker_thread.pas
optdiff skip BUILD-FAIL: c_obj_fnptr_a.c case_sensitive_unit.pas kwarg_overload_unit.pas lib_synapse_tls_loopback.pas shadow_b.pas strhelperprobe.pas test_a_call_result_is_not_an_assignment_target_through_a_cast.pas test_a_used_unit_keeps_its_own_assertion_default.pas test_a_whole_array_destination_refuses_a_scalar.pas test_array_param_default_refused.pas test_file_element_type_fail.pas test_include_miss_fails.pas test_lazy_var_scope_fail.pas test_mgmt_operators_field_refused.pas test_pointer_alias_identity.pas test_scalar_member_fail.pas test_sealed_abstract_class_fail.pas test_shared_lib.c test_string_default_on_ordinal_param_fail.pas unit_end_shapes_b.pas
optdiff THREADSAFE-RETRY: test_halt_from_worker_thread.pas test_parallel_for_capture_aggr.pas test_parallel_for_capture_string.pas test_parallel_policy.pas test_parallel_policy_lang.pas test_thread_api_no_uses.pas test_threadsafe_heap_lock_release.pas test_threadsafe_i386_stress.pas
optdiff shard 0/12: pass=210 skip=24 diff=1

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

RESOLVED 2026-09-16 (frankS, Track O/A) — 84ccb6384

The parent's -O3 arm was the INLINER, not a second copy of the conversion rule.

ir.inc's float→int rewrite (wrap the RHS in the -204 Round intrinsic) lives in the AN_ASSIGN lowering. For a single-statement body Result := <float expr> the inliner takes shape 1, which retains only the RHS expression and discards the AN_ASSIGN. Probed at the top of that arm: zero inlined assignments with a float RHS ever reach it. Nothing was wrong with the conversion — there was no assignment left to convert.

What isolates it: out-of-line at -O3 is correct, and the same body written to a local rather than Result is correct at -O3 (viaResult=-858993459 vs viaLocal=5). Shape 3 allocates a properly typed Result temp and stores through it, which is an AN_ASSIGN and does get the rewrite.

The fix reuses a precedent rather than adding a rule. The guard directly above already covered the mirror case — a conversion into a float result, the D2S bug — and its own comment states the general rule: "any RHS kind that is not ALREADY the result kind goes to shape 3". The implemented predicate was narrower than its own stated intent: keyed on the RESULT being float, so structurally blind to a float RHS landing in an integer result. Extended to the sibling, routing to TryRetainInlineStmtBody. Deliberately not a copy of the rounding rule — ir.inc's arm says two dozen places build an AN_ASSIGN and exactly one lowers it.

Measured, not assumed:

instrument result
PXXDBG=a.inline on the repro exactly ONE line moves: RetInt shape=1 → shape=3. Still RETAINED — a change of shape, not a lost inline.
same, across compiler/compiler.pas all 206 retentions byte-identical to pin v410 — the guard fires on nothing in the compiler's own source. The instrument is proven live by the one repro line that does move.
positive control: pin v410, byte-identical sources still -858993459, viaResult=-858993459, got 4616977747989548237 want 5. This tree: 5 at -O0/-O1/-O2/-O3.

viaLocal=5 on both sides is the aiming check: the guard did not blanket-disable the expression path, it diverted one shape.

No new fixture. The optdiff sweep is the designed instrument for an opt-level disagreement and test_double_to_integer_lvalue_rounds.pas at -O3 going green IS the regression test; a bespoke Makefile row would duplicate it.

Self-host fixedpoint converged, gate.sh quick GREEN. Inert for anything building against $(PXX_STABLE) until the next pin.