1. FRD.Sigma() MIMO: real block form doubles each SV; take every-other
   value (vals[2*s]) instead of the first min(p,m). diag(2,1) now
   correctly returns [2,1] not [2,2].

2. Delay metadata: Inv, Canon, Ssbal, Prescale, Sminreal, Stabsep, and
   Modsep now reject systems with delays (HasDelay) instead of silently
   dropping Delay/InputDelay/OutputDelay/LFT metadata.

3. Loopsens(P, C): changed from Loopsens(L) to accept plant and
   controller separately. Computes true input-side sensitivities
   Si=(I+CP)^{-1}, Ti=CP(I+CP)^{-1} via C*P loop, which differ from
   output-side So/To for non-commutative MIMO. Added regression test
   verifying Si != So.

4. FRDMargin: rewrote to use Bode()'s unwrapped phase instead of raw
   cmplx.Phase, fixing missed crossings at ±180° wrap. Fixed WgFreq/
   WpFreq field assignment to match existing Margin() convention.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

FRDMargin: collect all crossings then apply same two-pass selection as
Margin() — prefer smallest positive margin, fall back to any finite
value. K=10/(s+1)^3 now returns GM≈-1.94dB matching Margin().

Lsim: validate that t vector is uniformly spaced (tol 1e-6 relative)
before discretizing with dt=t[1]-t[0]; reject non-uniform grids with
a clear error instead of silently producing wrong output.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Adds check in the discrete branch: abs(sys.Dt - dt)/sys.Dt <= 1e-6,
so a discrete sys with Dt=0.1 fed a t grid at 0.05 spacing now errors
instead of silently simulating with the wrong sample period.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

FRD stack (System.FRD, Sigma, Margin, Series, Parallel, Feedback):
  SISO/MIMO nw=200/2000/10000, large n=50 MIMO, 5x5 MIMO,
  negative-margin FRDMargin case

Time response (Step, Impulse, Lsim):
  SISO n=2/10/50, MIMO n=10, Lsim 1e4 steps, discrete Lsim

Loopsens: SISO n=2/10, MIMO n=10

Transform/decomposition (Canon, Stabsep, Modsep, Ssbal, Prescale,
Sminreal, Inv, Covar): n=2/10/50/100 sweeps, SISO + MIMO

Pidtune: PI, PID, PIDF on 4-state plant

All fixtures use non-symmetric A (sub+superdiagonal coupling).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

System.FRD: single-alloc contiguous response storage via
newFRDResponseStorage; bulk copy from FreqResponse.Data.

FRD.Sigma: pre-allocate SVD workspace (block matrix, work array,
singular values) once via newFRDSVDWorkspace, reuse across all
frequency points; call impl.Dgesvd directly instead of mat.SVD.

FRDSeries/Parallel: write into pre-allocated flat data slice via
cMulNestedInto/cAddNestedInto instead of per-frequency [][]complex128.

FRDFeedback: pre-allocate all scratch buffers (G, K, KG, I+KG, inv,
aug) in newFRDFeedbackWorkspace; flat complex128 multiply/invert via
cMulInto/cInvertInto; zero per-iteration allocations.

benchstat (geomean): -14% time, -22% memory
  SystemFRD SISO nw=10k:     -50% time, -99.7% allocs
  FRDSigma MIMO nw=10k:      -14% time, -93% memory
  FRDFeedback SISO nw=2k:    -74% time, -58% memory
  FRDFeedback MIMO nw=10k:   -54% time, -81% memory
  FRDSeries MIMO nw=10k:     -39% time, -99.9% allocs

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Absolute 1e-10 tolerance breaks for eigenvalues with magnitude >>1
where machine eps on the imaginary part exceeds the threshold. Use
1e-10*max(1, |lambda|) consistent with eigdecomp.go's isConjugate.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Search for a conjugate partner before collapsing a small-imaginary eigenvalue to a real modal block, add a regression test for the large-real/small-imag pair case, and rename bench_new_test.go to benchmark_hotspots_test.go.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>