rtx-fsi: decouple the retry threshold from the widened stall acceptance

The first STALLX draft used one threshold for both, which silently
disabled the history-reset retry for the widened band: a stall in
[5x, STALLX x) was accepted unretried where runs 2/3 retried it —
caught by trajectory divergence at t = 5.7 (runs 2/3 digit-identical
there; the run-5 variant killed at 2.3 h, artifacts preserved as
*_unretried_variant_*). Now `retry_at` stays the old 5x-or-increment
window (retry semantics bit-identical to the pre-knob code at any
STALLX), and `stall_accept` widens only the post-retry acceptance.
Default STALLX=5 keeps both thresholds equal — committed defaults
unchanged.

Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Lnyrw33Lu6rUhW42E9KHwq

crates/specialized/rtx-fsi/tests/fsi2_harness/march.rs

@@ -451,13 +451,20 @@ pub fn run_march(case: BenchmarkCase, config: &MarchConfig) -> MarchResult {
             .sum::<f64>()
             .sqrt();
         let tol_step = tol_floor.max(rtol * increment);
-        // Acceptance beyond the tolerance: `stall_accept` x the
-        // tolerance (noise bouncing over a well-predicted step; see the
-        // config field for why FSI3's cycle needs a wider multiple than
-        // the default 5) or an order below the step's own increment
-        // (the rare violent step near peak motion — the s = 1 FSI2 run
-        // died at residual = 9% of its increment). Counted as stalls
-        // and bounded by the caller.
+        // Two thresholds, deliberately decoupled (the first STALLX
+        // draft used one and silently DISABLED the history-reset retry
+        // for the widened band — caught by trajectory divergence at
+        // t = 5.7 where runs 2/3 were digit-identical):
+        // `retry_at` is the OLD acceptance (5x the tolerance, or an
+        // order below the step's own increment — the rare violent step
+        // near peak motion; the s = 1 FSI2 run died at residual = 9% of
+        // its increment): any stall at or above it still gets the
+        // measured-valuable history reset + one retry from the
+        // predictor. `acceptable` (stall_accept x, default the same 5x
+        // — bit-identical) widens only the POST-RETRY acceptance, so a
+        // developed-cycle bistable stall the retry cannot fix is
+        // accepted and counted instead of fatal.
+        let retry_at = (5.0 * tol_step).max(0.1 * increment);
         let acceptable = (stall_accept * tol_step).max(0.1 * increment);
         let mut outcome = if let Some(iqn) = iqn.as_mut() {
             iqn.set_tolerance(tol_step).unwrap();
@@ -481,7 +488,7 @@ pub fn run_march(case: BenchmarkCase, config: &MarchConfig) -> MarchResult {
                 e,
                 rtx_fsi::FsiError::CouplingNotConverged { residual, .. }
                 | rtx_fsi::FsiError::CouplingDiverged { residual, .. }
-                if *residual >= acceptable
+                if *residual >= retry_at
             );
             if recoverable {
                 iqn_ref.reset_history();