GeezerVison

)

…upervisor 16:32:13Z.

amend-4 d4210b3 aligned (3-gate) fixture with empirically-confirmed
JIT smoke pattern (add+mul+fib force_compile + invoke). It did NOT fix
the script's auto-diagnostic over-claim that supervisor 16:32:13Z called
out: "counter == 0 (in stress fixture) ≠ site re-selection if JIT smoke
build counter > 0".

Diagnostic now correctly states: counter == 0 in (3-gate) BAD-PATH BUILD
indicates fixture-mismatch only, NOT site-re-selection. Site 600
traversal is independently confirmed by JIT smoke (Step 2) counter delta
> 0 in default build per testkeeper 16:31:45Z empirical evidence.

Self-attribution: amend-4 d4210b3 committed at 16:33:10Z, ~57s after
supervisor 16:32:13Z dispatch — same chat-recheck-race-window pattern
pythia python#381 16:32:05Z just flagged. Pre-emptive author-intuition commit
addressed the FIXTURE half of supervisor's request but missed the
DIAGNOSTIC half. amend-5 closes the gap. Two-action races (fixture +
diagnostic) decompose to single-action races at scale; this is signal
that pre-emptive commits should wait for explicit dispatch even when
direction is clear.

…_CMAKE_FLAGS.

Critical script bug discovered via testkeeper 17:03:02Z (3-gate)
diagnostic: build_phoenix.sh hard-reset EXTRA_CMAKE_FLAGS="" on every
invocation, silently dropping caller-supplied -DJIT_TEST_* flags. This
masked a series of c22b-mech amends — gate_phoenix.sh Step 1a-4 (3-gate)
substep set EXTRA_CMAKE_FLAGS=" -DJIT_TEST_VARIADIC_BRIDGE=1
-DJIT_TEST_EXERCISE=1" expecting it to reach cmake; build_phoenix.sh
overwrote with empty string before invoking cmake.

Result: bad-path build was actually NORMAL --pydebug build (no
JIT_TEST_VARIADIC_BRIDGE), so shadow path dispatched all args, so
JIT_CHECK count comparison passed silently. The "Immortal Object refcnt
underflow" runtime corruption testkeeper observed is a SEPARATE issue
in shadow-emit + linked-inputs interaction (likely def-use chain
corruption when shadow with allocateLinkedInput is deleted), NOT the
JIT_TEST_VARIADIC_BRIDGE bad-path firing.

Fix: EXTRA_CMAKE_FLAGS="${EXTRA_CMAKE_FLAGS:-}" preserves external value;
ASAN flag now appends rather than overwrites.

Cumulative c22b-mech: f47e4bf + 6944618 + 9ba37f8 + b16dff7
+ d4210b3 + 303eb0c + this commit. Net +1 LOC functional change
(EXTRA_CMAKE_FLAGS init), +5 LOC documentation.

Self-attribution: 4 amend cycles in this session would have been avoided
if I'd grep'd for EXTRA_CMAKE_FLAGS handling in build_phoenix.sh BEFORE
amend-3. Pre-emptive commit pattern (pythia python#381) compounds when the
fix is correct-but-untested by the agent.

Next: testkeeper re-run on this HEAD with JIT_VARIADIC_BAD_PATH_VERIFY=1;
expected outcome — flag actually reaches cmake, bad-path #ifdef compiles
in, shadow has count mismatch, JIT_CHECK fires (criterion 3 NEGATIVE
TEST gate-verified). Separately, shadow-emit + linked-inputs corruption
is a c22b-mech architecture issue requiring theologian review before
any push.