Algorithm A1 Synthetic anomaly injection
Require: Test values $X\in {\mathbb{R}}^{N\times F}$; train+val values $X_{tv}$ and alive mask $M_{tv}$; alive mask M; period $T_p$; frequency $freq$; severity $sev$; correlation threshold $\gamma$; max partners $p_{\max }$; cooldown parameter C
Ensure: Perturbed series $\tilde{X}$, window-level labels y
1:	$\tilde{X}\leftarrow X$; $O\leftarrow$ False                          ▹occupancy mask
2:	$n_f\leftarrow \mathrm{Q}99\left(X_{tv}[M_{tv}=1]\right)-\mathrm{Q}1\left(X_{tv}[M_{tv}=1]\right)$ for each feature f  ▹ global per-feature scale, train+val only
3:	$R\leftarrow$ pairwise absolute correlation matrix of X over non-constant features
4:	$P\leftarrow$ features with strong lag-$T_p$ autocorrelation; for each $f\in P$, locate local peaks/troughs (donor candidates)
5:	$k\leftarrow \lfloor freq·N/3\rfloor$; set target counts $\{k_{\mathrm{coll}},k_{\mathrm{ctx}},k_{\mathrm{pt}}\}$ per type
6:	$L_c\leftarrow 12$ (collective length), $L_x\leftarrow 3$ (contextual length)
7:	function Partners(f)
8:	return up to $p_{\max }$ features g with $R_{f,g}\ge \gamma$, ranked by correlation
9:	end function
10:	function TryCollective
11:	Sample start s, feature f; let $e\leftarrow s+L_c$
12:	if region $[s,e]$ free and f alive then
13:	Cycle mode $\in \{\mathrm{shift},\mathrm{variance}\mathrm{burst},\mathrm{frozen}\}$ (round-robin, no repeats until all used)
14:	Perturb $\tilde{X}[s:e,f]$ by magnitude $sev·n_f$ per mode; repeat on each feature in Partners(f) using its own $n_g$
15:	$y[s:e]\leftarrow 1$; mark $[s-C,e+C]$ occupied; return success
16:	end if
17:	end function
18:	function TryContextual
19:	Draw next unused (feature $f\in P$, extremum c) pair; let $[s,e]\leftarrow [c-1,c+2]$
20:	if region free and a same-feature donor extremum d of opposite phase exists with $|d-c|\ge 2L_x$ and correlation$(\tilde{X}[s:e,f],X[d-1:d+2,f])\le 0.85$ then
21:	Replace $\tilde{X}[s:e,f]$ with donor window $X[d-1:d+2,f]$; repeat for each partner in Partners(f) using its matching-phase donor
22:	$y[s:e]\leftarrow 1$; mark $[s-C,e+C]$ occupied; return success
23:	end if
24:	end function
25:	function TryPoint
26:	Sample timestamp t, feature f
27:	if $M[t,f]=1$ and t free then
28:	Cycle mode $\in \{\mathrm{spike},\mathrm{drop}\}$ (round-robin); set $\tilde{X}[t,f]\leftarrow \tilde{X}[t-1,f]\pm sev·n_f$; repeat on each partner in Partners(f)
29:	$y\left[t\right]\leftarrow 1$; mark $[t-C,t+C]$ occupied; return success
30:	end if
31:	end function
32:	Round-robin scheduling: while any active type is below its target count and below a failure-streak budget, cycle through {collective, contextual, point} in turn, attempting one placement per active type per pass; deactivate a type once its target count is reached or its failure streak is exceeded
33:	Assign each raw anomalous timestep as the positive label of its corresponding window; under this convention, window-level and point-level labels coincide one-to-one
34:	return $\tilde{X},y$