tr_tst2 baseline: backfill AJRFT after eq_tst2 drift fix

[k-yoshimi]

Summary

test_run/baselines/tr_tst2/metrics.json is missing the AJRFT scalar key. After commit 24b1b12f ("test+tr: backfill AJRFT in regression dump") + the libtrapi.so Python wrapper update from PR #187, libtrapi.so emits 14 scalars (including AJRFT) but the tst2 baseline still has 13. Whenever eq_tst2 is fixed and tr_tst2 runs in CI, the Layer-1 equivalence test will fail with:

compare_metrics FAIL: scalars.AJRFT: missing

Why this wasn't fixed in the original AJRFT PR

tr_tst2 depends on eq_tst2, which has its own pre-existing ~3e-9 physics drift on the current develop branch (eq_tst2 regression on clavius 2026-05-10 — likely upstream eq changes since the baseline was last regenerated). Regenerating the tr_tst2 baseline requires either:

• Resolving the eq_tst2 drift first, OR
• Manually patching just the AJRFT: 0.0 line (risky — assumes the actual value is 0; in-house code review on commit 24b1b12f explicitly cautioned against this)

The choice was made to leave tst2 with the existing stale baseline (which is the status quo — the test SKIPs locally due to missing eqdata, hiding the issue) and address it as a separate concern.

Acceptance

1. Resolve the eq_tst2 drift (separate work — investigate which develop change introduced the ~3e-9 shift; eq physics or eq baselines stale?)
2. Regenerate the eq_tst2 baseline on a Linux box with current eq (see \$CLAUDE_MEMORY_DIR/memory/reference_clavius_baseline_regen.md)
3. Then regenerate the tr_tst2 baseline (will include AJRFT automatically via the fix in 24b1b12f)
4. Verify pytest python/trlib/tests/test_equivalence.py --forked --timeout=120 --timeout-method=signal -v reports both test_iter01 AND test_tst2 PASSED at 1e-10 — no SKIP

Related

• Original AJRFT backfill: commit 24b1b12f (chore/pre-push-hook-worktree-compat)
• PR tr+totlib: physical EXTERNAL_DRIVEN_I scalar (L-7b-i) #187 (L-7b-i EXTERNAL_DRIVEN_I) — added AJRFT to C ABI but missed the regress paths
• PR chore(pre-push): make hook + CLAUDE.md instructions worktree-compatible #176 addendum (PNBTOT registry extension) — discovered + tracked this gap
• Memory: feedback_scalar_field_triangle.md, feedback_equivalence_must_pass.md

🤖 Filed via post-push retrospective on 2026-05-11.

[k-yoshimi]

Investigation update (2026-05-12, via attempted PR #196):

The drift was confirmed at clavius gfortran 13.3.0 — eq_tst2 baseline regen produces values that pass at 1e-10 against clavius standalone eq.x AND libeqapi.so, but the CI environment (Ubuntu 24.04 default gfortran 13.2.0) produces shifted values (~3e-9 in profile arrays only, not scalars).

Initial hypothesis was that a98d66ec (fix(eq): defensive zero-init sweep #130) caused the drift via heap-reuse correctness fix. Investigation showed the drift is dominated by compiler-version difference (gfortran 13.2 vs 13.3), not zero-init. CI's output exactly matches the OLD baseline (rel_err = 0), suggesting 9614009's regen was on gfortran 13.2.

Resolving #190 cleanly requires aligning the baseline-regen environment with CI's environment. Tracked in #197 ("CI: baseline regen env mismatch — clavius gfortran 13.3 vs CI gfortran 13.2 produces ~3e-9 drift"). PR #196 closed without merging; the xfail decorator in trlib (added by PR #195) stays in place until #197 unblocks #190.