iter 17 — sibling-stage mechanics: a one-keystroke crash found and fixed

BREAKS → DEFENDED → CLOSED 2026-07-10 · the last §B stage directly attacked · duplicate-name slate crashed raw, now deduped-with-warning · clean 0/3 (reset)

Claim under attack (assumed true only to destroy it)

The sibling stage (Finish — mutual-duplicate detection within a slate) is operationally sound: (i) a slate containing a duplicate kernel name — a one-keystroke operator error in the terminal phase's --names lists — is handled sanely; (ii) sibling fields are well-formed on real multi-kernel slates; (iii) the PENDING-SIBLING guard degrades a would-be PASS with unmeasurable siblings.

Result matrix

Falsifier	Result
F1 — duplicate-name slate	FIRED, exactly as code-reading predicted: `candidate_matrix`'s name-keyed dict collapses to ONE column while `names` keeps two → raw `IndexError: index 1 is out of bounds for axis 1 with size 1` deep in numpy. A slate typo kills the whole run.
F2 — sibling fields on a real 8-kernel slate	8/8 well-formed (`sibling_worst_rho`, `sibling_worst_loo_r2`, `stages.sibling`)
F3 — PENDING-SIBLING guard (conditional)	CONDITIONAL-UNTESTABLE on live kernels: no PASS-verdict kernel in the discovery batch (2 BAN · 3 WATCH · 3 PENDING on the 2-cell smoke grid); the guard's logic remains pinned by the declaration self-test (PASS + no sibling measurement → PENDING)

DEFENSE — slate hygiene at both verdict APIs

aggregate_worst_cell / aggregate_worst_regime:
+   names = list(dict.fromkeys(names))          # dedup, order-preserving
+   stderr: "WARNING: N duplicate candidate name(s) removed from the slate"

Re-run: the dup slate completes with exactly one well-formed entry (verdict WATCH for sevcik_fd, all hardened fields present); §B declaration self-test PASS.

Bonus intel — live-kernel verdict landscape (8 kernels, 2-cell smoke)

bartels_rank_vn_ratio BAN (shipped twin) · semivar_signed_ratio BAN · sevcik_fd / updown_cjs / anderson_darling_a2 WATCH · edge_spread_bps / hvg_clustering / petrosian_fd PENDING (all three with iter-13 failure records naming the cause). Useful triage for the terminal phase.

Evidence

File	What
a17_sibling_stage_attack.py	The attack — falsifiers pre-declared in-file.
a17_results.json · a17_batch_verdict.json · a17_dup_verdict.json	Falsifier evaluation + the probe's own batch and (post-fix) dup-slate verdicts.
a17_stderr.log · a17_stderr_postfix.log	Pre-fix (raw IndexError) and post-fix (dedup warning) runs under the 2 CPU / 2 GB scope.
regression_tests/test_a17_sibling_stage.sh	Permanent regression test — re-runs the attack, asserts SURVIVES.

Consequence

The 14th finding: an input-hygiene crash on the exact path the terminal phase will drive with hand-written --names lists. Every §B stage has now been directly attacked. Clean rounds: reset to 0 of 3.

iter 17 · sibling-stage mechanics · verdict BREAKS → DEFENDED → CLOSED · one atomic commit: iter 17: sibling-stage mechanics — BREAKS→DEFENDED · append-only.